@mlx-bun/inference 0.0.0-stage → 1.0.0
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/LICENSE +21 -0
- package/README.md +1475 -2
- package/THIRD_PARTY_NOTICES.md +3404 -0
- package/dist/native/libmlx_bun_ane.dylib +0 -0
- package/dist/native/libmlx_bun_expert_io.dylib +0 -0
- package/dist/native/mlx-bun-frame-extract +0 -0
- package/examples/fit-model.ts +25 -0
- package/examples/qwen3-forward.ts +42 -0
- package/examples/qwen3-generate.ts +28 -0
- package/examples/trellis-expand.ts +23 -0
- package/package.json +129 -4
- package/src/adapters/available.ts +110 -0
- package/src/adapters/index.ts +3 -0
- package/src/adapters/manager.ts +316 -0
- package/src/artifacts/auxiliary-files.ts +45 -0
- package/src/artifacts/config-dialects.ts +94 -0
- package/src/artifacts/config.ts +341 -0
- package/src/artifacts/draft-vocabulary.ts +30 -0
- package/src/artifacts/expert-offload-build.ts +75 -0
- package/src/artifacts/expert-offload.ts +89 -0
- package/src/artifacts/experts/index.ts +3 -0
- package/src/artifacts/experts/io.ts +386 -0
- package/src/artifacts/experts/residency.ts +1134 -0
- package/src/artifacts/experts/usage.ts +415 -0
- package/src/artifacts/fingerprint.ts +44 -0
- package/src/artifacts/glm52/container.ts +305 -0
- package/src/artifacts/glm52/expert-layout.ts +249 -0
- package/src/artifacts/glm52/index.ts +7 -0
- package/src/artifacts/glm52/kernel-layout.ts +124 -0
- package/src/artifacts/glm52/layout.ts +152 -0
- package/src/artifacts/glm52/memory.ts +500 -0
- package/src/artifacts/glm52/resident-weights.ts +135 -0
- package/src/artifacts/glm52/weights.ts +503 -0
- package/src/artifacts/glm52-config.ts +262 -0
- package/src/artifacts/identity.ts +94 -0
- package/src/artifacts/index.ts +15 -0
- package/src/artifacts/mmap.ts +90 -0
- package/src/artifacts/safetensors-writer.ts +249 -0
- package/src/artifacts/safetensors.ts +164 -0
- package/src/artifacts/turboquant-fold.ts +42 -0
- package/src/artifacts/weight-audit.ts +43 -0
- package/src/artifacts/weights.ts +255 -0
- package/src/contracts/index.ts +2 -0
- package/src/contracts/mlx/attention.ts +8 -0
- package/src/contracts/mlx/cache.ts +510 -0
- package/src/contracts/mlx/checkpoint.ts +10 -0
- package/src/contracts/mlx/draft-checkpoint.ts +14 -0
- package/src/contracts/mlx/draft-projection.ts +27 -0
- package/src/contracts/mlx/draft-target.ts +84 -0
- package/src/contracts/mlx/drafter.ts +176 -0
- package/src/contracts/mlx/forward-work.ts +23 -0
- package/src/contracts/mlx/glm52-weights.ts +25 -0
- package/src/contracts/mlx/graph.ts +309 -0
- package/src/contracts/mlx/head.ts +6 -0
- package/src/contracts/mlx/index.ts +15 -0
- package/src/contracts/mlx/kv-codec.ts +9 -0
- package/src/contracts/mlx/media.ts +113 -0
- package/src/contracts/mlx/positions.ts +16 -0
- package/src/contracts/mlx/row-table.ts +2 -0
- package/src/contracts/mlx/trainable.ts +103 -0
- package/src/contracts/mlx/whisper.ts +96 -0
- package/src/contracts/portable/composition.ts +106 -0
- package/src/contracts/portable/denoising.ts +17 -0
- package/src/contracts/portable/execution.ts +79 -0
- package/src/contracts/portable/generation.ts +103 -0
- package/src/contracts/portable/graph.ts +80 -0
- package/src/contracts/portable/index.ts +15 -0
- package/src/contracts/portable/memory-plan.ts +26 -0
- package/src/contracts/portable/object-cache.ts +8 -0
- package/src/contracts/portable/prefill.ts +63 -0
- package/src/contracts/portable/prefix-cache.ts +33 -0
- package/src/contracts/portable/prompt.ts +2 -0
- package/src/contracts/portable/resources.ts +25 -0
- package/src/contracts/portable/rollback.ts +7 -0
- package/src/contracts/portable/sampling.ts +44 -0
- package/src/contracts/portable/scheduling.ts +54 -0
- package/src/contracts/portable/token-delivery.ts +4 -0
- package/src/contracts/portable/tsconfig.json +14 -0
- package/src/embeddings/index.ts +1 -0
- package/src/embeddings/text.ts +85 -0
- package/src/execution/admission.ts +59 -0
- package/src/execution/batch-group.ts +1240 -0
- package/src/execution/batch-types.ts +225 -0
- package/src/execution/cancellation.ts +38 -0
- package/src/execution/continuation-request.ts +45 -0
- package/src/execution/continuation-types.ts +10 -0
- package/src/execution/coordinator.ts +70 -0
- package/src/execution/denoising-group.ts +180 -0
- package/src/execution/engine.ts +56 -0
- package/src/execution/fill-group.ts +505 -0
- package/src/execution/fit.ts +184 -0
- package/src/execution/gateway-binding.ts +178 -0
- package/src/execution/grammar-group.ts +224 -0
- package/src/execution/index.ts +18 -0
- package/src/execution/methods.ts +75 -0
- package/src/execution/mixed-iteration.ts +96 -0
- package/src/execution/plan.ts +65 -0
- package/src/execution/prefill-cohort.ts +100 -0
- package/src/execution/prefill-rows.ts +263 -0
- package/src/execution/request-policy.ts +19 -0
- package/src/execution/row-sampling.ts +49 -0
- package/src/execution/scheduler.ts +92 -0
- package/src/execution/session.ts +247 -0
- package/src/execution/speculative-group.ts +797 -0
- package/src/execution/tasks.ts +44 -0
- package/src/generation/autoregressive.ts +840 -0
- package/src/generation/bindings/autoregressive.ts +104 -0
- package/src/generation/bindings/denoising.ts +60 -0
- package/src/generation/bindings/prefill-policy.ts +36 -0
- package/src/generation/checkpoint-identity.ts +60 -0
- package/src/generation/denoising.ts +85 -0
- package/src/generation/diffusion.ts +583 -0
- package/src/generation/fill/echo-index.ts +188 -0
- package/src/generation/fill/index.ts +3 -0
- package/src/generation/fill/proposal.ts +69 -0
- package/src/generation/fill/schema-rows.ts +506 -0
- package/src/generation/fill/session.ts +464 -0
- package/src/generation/fill/tool-boundary.ts +103 -0
- package/src/generation/fill-append.ts +58 -0
- package/src/generation/generate.ts +44 -0
- package/src/generation/grammar-step.ts +39 -0
- package/src/generation/index.ts +14 -0
- package/src/generation/phase-dispatch.ts +115 -0
- package/src/generation/prefill.ts +33 -0
- package/src/generation/result.ts +49 -0
- package/src/generation/scopes.ts +157 -0
- package/src/generation/speculative/acceptance.ts +41 -0
- package/src/generation/speculative/adaptive-draft-gate.ts +166 -0
- package/src/generation/speculative/bindings/assistant-rows.ts +102 -0
- package/src/generation/speculative/bindings/autoregressive-rows.ts +124 -0
- package/src/generation/speculative/bindings/binding.ts +106 -0
- package/src/generation/speculative/bindings/deepspec-rows.ts +33 -0
- package/src/generation/speculative/bindings/draft-target.ts +10 -0
- package/src/generation/speculative/bindings/glm52-mtp-rows.ts +128 -0
- package/src/generation/speculative/bindings/projected-context-rows.ts +78 -0
- package/src/generation/speculative/bindings/projected-draft-rows.ts +95 -0
- package/src/generation/speculative/bindings/qwen-mtp-rows.ts +336 -0
- package/src/generation/speculative/bindings/two-model-rows.ts +63 -0
- package/src/generation/speculative/cache-identity.ts +16 -0
- package/src/generation/speculative/draft-kind.ts +149 -0
- package/src/generation/speculative/draft-policy.ts +41 -0
- package/src/generation/speculative/draft-registry.ts +118 -0
- package/src/generation/speculative/draft-vocab.ts +110 -0
- package/src/generation/speculative/index.ts +17 -0
- package/src/generation/speculative/output.ts +49 -0
- package/src/generation/speculative/round.ts +307 -0
- package/src/generation/speculative/source.ts +144 -0
- package/src/generation/speculative/sources/assistant-source.ts +64 -0
- package/src/generation/speculative/sources/deepspec-source.ts +70 -0
- package/src/generation/speculative/sources/dflash-source.ts +84 -0
- package/src/generation/speculative/sources/dflash2-source.ts +86 -0
- package/src/generation/speculative/sources/glm52-mtp-source.ts +66 -0
- package/src/generation/speculative/sources/ngram-source.ts +165 -0
- package/src/generation/speculative/sources/qwen-mtp-source.ts +154 -0
- package/src/generation/speculative/sources/two-model.ts +75 -0
- package/src/generation/types.ts +194 -0
- package/src/index.ts +20 -0
- package/src/input/audio/audiotoolbox.ts +112 -0
- package/src/input/audio/decode.ts +133 -0
- package/src/input/audio/features.ts +202 -0
- package/src/input/audio/index.ts +6 -0
- package/src/input/audio/transcode.ts +67 -0
- package/src/input/audio/whisper-mel.ts +185 -0
- package/src/input/audio/whisper-tokenizer.ts +238 -0
- package/src/input/chat-template.ts +389 -0
- package/src/input/index.ts +6 -0
- package/src/input/media-fetch.ts +439 -0
- package/src/input/packed-mixed.ts +43 -0
- package/src/input/special-tokens.ts +75 -0
- package/src/input/token-groups.ts +49 -0
- package/src/input/tokenizer.ts +336 -0
- package/src/input/tool-call.ts +689 -0
- package/src/input/vision/index.ts +3 -0
- package/src/input/vision/preprocess.ts +247 -0
- package/src/input/vision/prompt.ts +398 -0
- package/src/input/vision/qwen3vl-preprocess.ts +409 -0
- package/src/input/vision/qwen3vl-prompt.ts +237 -0
- package/src/input/vision/video-frames.ts +100 -0
- package/src/kernels/ane/fill.ts +101 -0
- package/src/kernels/ane/linear.ts +416 -0
- package/src/kernels/attention/flash.ts +694 -0
- package/src/kernels/attention/kvq4-decode.ts +155 -0
- package/src/kernels/attention/masks.ts +40 -0
- package/src/kernels/attention/multi-query.ts +193 -0
- package/src/kernels/attention/paged.ts +130 -0
- package/src/kernels/delta/gated.ts +306 -0
- package/src/kernels/delta/index.ts +3 -0
- package/src/kernels/delta/silu-rms-heads.ts +69 -0
- package/src/kernels/delta/state.ts +74 -0
- package/src/kernels/gated-rms-norm.ts +64 -0
- package/src/kernels/glm52/layout.ts +37 -0
- package/src/kernels/glm52/streamed-experts.ts +360 -0
- package/src/kernels/logits.ts +38 -0
- package/src/kernels/quantization/affine-row.ts +154 -0
- package/src/kernels/quantization/affine-rows.ts +36 -0
- package/src/kernels/quantization/affine-split-mma.ts +144 -0
- package/src/kernels/quantization/affine3-mma.ts +117 -0
- package/src/kernels/quantization/affine3-rows.ts +103 -0
- package/src/kernels/quantization/affine4-mma.ts +113 -0
- package/src/kernels/quantization/affine4-rows.ts +102 -0
- package/src/kernels/residual-rms-norm.ts +65 -0
- package/src/kernels/sampling/index.ts +2 -0
- package/src/kernels/sampling/normalized-argmax.ts +116 -0
- package/src/kernels/sampling/token-bitmask.ts +52 -0
- package/src/kernels/softcap.ts +35 -0
- package/src/kernels/trellis/codebook.ts +123 -0
- package/src/kernels/trellis/down-codebook.ts +184 -0
- package/src/kernels/trellis/down-factored.ts +180 -0
- package/src/kernels/trellis/down-k3-interleaved-factored.ts +146 -0
- package/src/kernels/trellis/down-mma-wide.ts +195 -0
- package/src/kernels/trellis/down-mma.ts +145 -0
- package/src/kernels/trellis/expand.ts +42 -0
- package/src/kernels/trellis/gate-up-codebook-mma.ts +141 -0
- package/src/kernels/trellis/gate-up-codebook.ts +137 -0
- package/src/kernels/trellis/gate-up-factored.ts +197 -0
- package/src/kernels/trellis/gate-up-mma-wide.ts +147 -0
- package/src/kernels/trellis/gate-up-mma.ts +148 -0
- package/src/kernels/trellis/gate-up.ts +248 -0
- package/src/kernels/trellis/geometry.ts +22 -0
- package/src/kernels/trellis/index.ts +103 -0
- package/src/kernels/trellis/launch.ts +12 -0
- package/src/kernels/trellis/mixed-gate-up-factored.ts +119 -0
- package/src/kernels/trellis/mixed-gate-up.ts +146 -0
- package/src/kernels/trellis/qualified.ts +34 -0
- package/src/kernels/trellis/reduce.ts +150 -0
- package/src/kernels/trellis/scatter.ts +407 -0
- package/src/kernels/trellis/splitk-prefill.ts +152 -0
- package/src/kernels/trellis/tiled-prefill.ts +131 -0
- package/src/kernels/trellis/vector-expand.ts +58 -0
- package/src/kernels/trellis/wide-prefill.ts +89 -0
- package/src/kernels/turboquant/decode.ts +82 -0
- package/src/kernels/turboquant/index.ts +5 -0
- package/src/kernels/turboquant/inverse.ts +122 -0
- package/src/kernels/turboquant/ops.ts +860 -0
- package/src/kernels/turboquant/signs.ts +130 -0
- package/src/kernels/turboquant/tables.ts +124 -0
- package/src/layers/affine-verify-linear.ts +75 -0
- package/src/layers/ane-prefill-split.ts +283 -0
- package/src/layers/compiled-normalization.ts +41 -0
- package/src/layers/delta-heads.ts +23 -0
- package/src/layers/dense-embedding.ts +29 -0
- package/src/layers/dense-linear.ts +32 -0
- package/src/layers/dequantized-linear.ts +72 -0
- package/src/layers/folded-attention.ts +42 -0
- package/src/layers/geglu.ts +130 -0
- package/src/layers/helpers.ts +7 -0
- package/src/layers/index.ts +36 -0
- package/src/layers/kv4-decode-attention.ts +15 -0
- package/src/layers/layer-norm.ts +24 -0
- package/src/layers/loading.ts +92 -0
- package/src/layers/lora.ts +150 -0
- package/src/layers/normalization.ts +9 -0
- package/src/layers/prefix-layout.ts +33 -0
- package/src/layers/quantized-append-attention.ts +32 -0
- package/src/layers/quantized-attention.ts +353 -0
- package/src/layers/quantized-embedding.ts +39 -0
- package/src/layers/quantized-linear.ts +137 -0
- package/src/layers/quantized-switch-linear.ts +37 -0
- package/src/layers/qwen-append.ts +12 -0
- package/src/layers/qwen-mrope.ts +238 -0
- package/src/layers/rope-2d.ts +110 -0
- package/src/layers/rope.ts +389 -0
- package/src/layers/swiglu.ts +23 -0
- package/src/layers/trellis-codebook-down.ts +52 -0
- package/src/layers/trellis-codebook-gate-up.ts +42 -0
- package/src/layers/trellis-down.ts +117 -0
- package/src/layers/trellis-gate-up.ts +71 -0
- package/src/layers/trellis-linear.ts +211 -0
- package/src/layers/trellis-prompt.ts +102 -0
- package/src/layers/trellis-reduce-gate-up.ts +73 -0
- package/src/layers/trellis-scatter-down.ts +98 -0
- package/src/models/ane-install.ts +74 -0
- package/src/models/audio/conformer.ts +656 -0
- package/src/models/cache-composition.ts +44 -0
- package/src/models/capabilities.ts +89 -0
- package/src/models/chat-template.ts +10 -0
- package/src/models/composition.ts +107 -0
- package/src/models/diffusion-gemma/experts.ts +59 -0
- package/src/models/diffusion-gemma/family.ts +21 -0
- package/src/models/diffusion-gemma/model.ts +900 -0
- package/src/models/diffusion-gemma/vision.ts +549 -0
- package/src/models/drafter-loaders.ts +49 -0
- package/src/models/drafters.ts +39 -0
- package/src/models/factory.ts +351 -0
- package/src/models/families.ts +70 -0
- package/src/models/family.ts +125 -0
- package/src/models/gemma4/assistant-target.ts +46 -0
- package/src/models/gemma4/assistant.ts +327 -0
- package/src/models/gemma4/blocks.ts +755 -0
- package/src/models/gemma4/compiled-decode.ts +452 -0
- package/src/models/gemma4/family.ts +39 -0
- package/src/models/gemma4/gemma4-12b.ts +677 -0
- package/src/models/gemma4/gemma4-26b.ts +544 -0
- package/src/models/gemma4/gemma4-e4b.ts +599 -0
- package/src/models/gemma4/graphs.ts +147 -0
- package/src/models/gemma4/media.ts +88 -0
- package/src/models/gemma4/model.ts +860 -0
- package/src/models/gemma4/serving-blocks.ts +72 -0
- package/src/models/gemma4/trainable.ts +186 -0
- package/src/models/gemma4/written-blocks.ts +232 -0
- package/src/models/gemma4/written-forward.ts +159 -0
- package/src/models/gemma4/written-input.ts +29 -0
- package/src/models/glm52/atlas.ts +507 -0
- package/src/models/glm52/coupling.ts +388 -0
- package/src/models/glm52/dsa.ts +601 -0
- package/src/models/glm52/family.ts +20 -0
- package/src/models/glm52/mla.ts +812 -0
- package/src/models/glm52/model.ts +1127 -0
- package/src/models/glm52/moe.ts +223 -0
- package/src/models/glm52/mtp.ts +60 -0
- package/src/models/glm52/pilot.ts +537 -0
- package/src/models/glm52/reference.ts +445 -0
- package/src/models/glm52/residency.ts +557 -0
- package/src/models/glm52/streamed-experts.ts +447 -0
- package/src/models/graph.ts +59 -0
- package/src/models/implementation.ts +230 -0
- package/src/models/index.ts +15 -0
- package/src/models/media-input.ts +15 -0
- package/src/models/memory-plan.ts +39 -0
- package/src/models/minicpm5/family.ts +29 -0
- package/src/models/minicpm5/model.ts +394 -0
- package/src/models/minicpm5/trainable.ts +90 -0
- package/src/models/profile.ts +337 -0
- package/src/models/qwen/blocks.ts +658 -0
- package/src/models/qwen/checkpoint.ts +190 -0
- package/src/models/qwen/families.ts +57 -0
- package/src/models/qwen/media-input.ts +161 -0
- package/src/models/qwen/mtp.ts +101 -0
- package/src/models/qwen/qwen3-moe.ts +355 -0
- package/src/models/qwen/qwen3.ts +323 -0
- package/src/models/qwen/qwen38-27b-trellis-m1max.ts +1532 -0
- package/src/models/qwen/qwen38-27b-trellis-m4pro.ts +1092 -0
- package/src/models/qwen/qwen3_5.ts +433 -0
- package/src/models/qwen/written-layers.ts +285 -0
- package/src/models/row-table.ts +1 -0
- package/src/models/runtime.ts +73 -0
- package/src/models/speculative/deepspec-artifact.ts +10 -0
- package/src/models/speculative/deepspec.ts +744 -0
- package/src/models/speculative/dflash.ts +553 -0
- package/src/models/speculative/dflash2-artifact.ts +9 -0
- package/src/models/speculative/dflash2-operations.ts +1 -0
- package/src/models/speculative/dflash2.ts +576 -0
- package/src/models/speculative/loader.ts +29 -0
- package/src/models/support.ts +79 -0
- package/src/models/universal/archs.ts +365 -0
- package/src/models/universal/blocks.ts +479 -0
- package/src/models/universal/dense.ts +255 -0
- package/src/models/universal/family.ts +17 -0
- package/src/models/vision/qwen3vl.ts +566 -0
- package/src/models/vision/siglip.ts +520 -0
- package/src/models/vision/unified.ts +176 -0
- package/src/models/whisper/family.ts +19 -0
- package/src/models/whisper/fast.ts +734 -0
- package/src/models/whisper/model.ts +430 -0
- package/src/runtime/compiled-trace.ts +16 -0
- package/src/runtime/config.ts +77 -0
- package/src/runtime/expert-trace.ts +71 -0
- package/src/runtime/graph-forward.ts +65 -0
- package/src/runtime/integers.ts +6 -0
- package/src/runtime/native.ts +16 -0
- package/src/runtime/resources.ts +53 -0
- package/src/runtime/row-operations.ts +42 -0
- package/src/runtime/trace.ts +161 -0
- package/src/sampling/curve.ts +122 -0
- package/src/sampling/draft.ts +55 -0
- package/src/sampling/extras.ts +45 -0
- package/src/sampling/filters.ts +134 -0
- package/src/sampling/grammar.ts +630 -0
- package/src/sampling/hlg.ts +290 -0
- package/src/sampling/index.ts +9 -0
- package/src/sampling/processors.ts +142 -0
- package/src/sampling/sampler.ts +120 -0
- package/src/sampling/step.ts +343 -0
- package/src/sampling/types.ts +161 -0
- package/src/sampling/window-filters.ts +46 -0
- package/src/scoring/decisions.ts +96 -0
- package/src/scoring/full-sequence.ts +103 -0
- package/src/scoring/index.ts +9 -0
- package/src/scoring/kl.ts +29 -0
- package/src/scoring/perplexity.ts +165 -0
- package/src/scoring/response-logps.ts +65 -0
- package/src/state/affine-attention.ts +180 -0
- package/src/state/attachment.ts +68 -0
- package/src/state/attention-read.ts +215 -0
- package/src/state/batched-kv.ts +294 -0
- package/src/state/batched-mask.ts +445 -0
- package/src/state/batched-quant.ts +389 -0
- package/src/state/batched-quantized-kv.ts +209 -0
- package/src/state/batched-rotating-quant.ts +457 -0
- package/src/state/batched-rotating-state.ts +149 -0
- package/src/state/batched-rotating.ts +430 -0
- package/src/state/batched-row-storage.ts +179 -0
- package/src/state/batched-ssm.ts +54 -0
- package/src/state/batched-turboquant-kv.ts +144 -0
- package/src/state/bf16-first-kv.ts +663 -0
- package/src/state/cache-service.ts +592 -0
- package/src/state/capabilities.ts +65 -0
- package/src/state/checkpoint.ts +46 -0
- package/src/state/continuation-persistence.ts +126 -0
- package/src/state/decode-step-inputs.ts +31 -0
- package/src/state/decoded-kv-donor.ts +12 -0
- package/src/state/delayed-quantized-kv.ts +141 -0
- package/src/state/delayed-rotating-quantized-kv.ts +178 -0
- package/src/state/delayed-turboquant-kv.ts +123 -0
- package/src/state/durability.ts +214 -0
- package/src/state/encoder-cache.ts +26 -0
- package/src/state/full-kv-row-append.ts +33 -0
- package/src/state/full-kv-row-donor.ts +54 -0
- package/src/state/full-prefill-padding.ts +56 -0
- package/src/state/full-prefill-row.ts +112 -0
- package/src/state/full-transitioning-kv-rows.ts +70 -0
- package/src/state/glm52-cache.ts +783 -0
- package/src/state/grouped-heads-kv4-cache.ts +26 -0
- package/src/state/index.ts +38 -0
- package/src/state/kv-attention-view.ts +77 -0
- package/src/state/kv-budget.ts +14 -0
- package/src/state/kv-scheme.ts +315 -0
- package/src/state/kv-storage.ts +57 -0
- package/src/state/kv-tensor-rows.ts +212 -0
- package/src/state/kv.ts +324 -0
- package/src/state/kv4-head256-cache.ts +46 -0
- package/src/state/latent-legos.ts +895 -0
- package/src/state/layout.ts +277 -0
- package/src/state/leases.ts +26 -0
- package/src/state/paged/cache.ts +689 -0
- package/src/state/paged/codec.ts +36 -0
- package/src/state/paged/rows.ts +223 -0
- package/src/state/persistence-types.ts +173 -0
- package/src/state/persistence-worker.ts +83 -0
- package/src/state/persistence.ts +1036 -0
- package/src/state/persistence.worker.d.ts +3 -0
- package/src/state/persistence.worker.js +264 -0
- package/src/state/prefix-cache.ts +685 -0
- package/src/state/quantized-kv.ts +369 -0
- package/src/state/quantized-tensor.ts +26 -0
- package/src/state/request-policy.ts +48 -0
- package/src/state/retention-policy.ts +37 -0
- package/src/state/rollback.ts +131 -0
- package/src/state/rotating-kv-donor.ts +35 -0
- package/src/state/rotating-kv-layout.ts +341 -0
- package/src/state/rotating-kv.ts +529 -0
- package/src/state/rotating-quantized-kv.ts +616 -0
- package/src/state/rotating-row-transaction.ts +35 -0
- package/src/state/rows.ts +104 -0
- package/src/state/softcap-attention.ts +118 -0
- package/src/state/speculative/glm52-mtp-state.ts +38 -0
- package/src/state/speculative/qwen-mtp-state.ts +34 -0
- package/src/state/speculative-rotating-kv.ts +113 -0
- package/src/state/ssd-cache.ts +434 -0
- package/src/state/ssm-prefill-padding.ts +39 -0
- package/src/state/ssm.ts +422 -0
- package/src/state/target-layout.ts +7 -0
- package/src/state/tiered-prefix-cache.ts +45 -0
- package/src/state/training-cache.ts +300 -0
- package/src/state/transitioning-kv-rows.ts +267 -0
- package/src/state/turboquant-attention.ts +97 -0
- package/src/state/turboquant-codec.ts +90 -0
- package/src/state/turboquant-kv.ts +442 -0
- package/src/state/views.ts +26 -0
- package/src/transcription/index.ts +3 -0
- package/src/transcription/silero-vad.ts +379 -0
- package/src/transcription/whisper/decode.ts +773 -0
- package/src/transcription/whisper/format.ts +74 -0
- package/src/transcription/whisper/index.ts +4 -0
- package/src/transcription/whisper/timing.ts +274 -0
- package/src/transcription/whisper/transcribe.ts +394 -0
- package/src/transcription/whisper/types.ts +22 -0
|
@@ -0,0 +1,1532 @@
|
|
|
1
|
+
import type { MlxArray } from "@mlx-bun/mlx/array";
|
|
2
|
+
import * as ops from "@mlx-bun/mlx/ops";
|
|
3
|
+
import type { ModelConfig } from "../../artifacts/config";
|
|
4
|
+
import type { Weights } from "../../artifacts/weights";
|
|
5
|
+
import type { Cache } from "../../contracts/mlx/cache";
|
|
6
|
+
import type { DraftProjectionKernels } from "../../contracts/mlx/draft-target";
|
|
7
|
+
import type { DraftProjection } from "../../contracts/mlx/draft-projection";
|
|
8
|
+
import type { AneProgramSpec } from "../../kernels/ane/linear";
|
|
9
|
+
import type { ResidualBasis, TargetView } from "../../contracts/mlx/draft-target";
|
|
10
|
+
import type { MlxDeclaredGraph, MlxTokenPhases, MlxLayerCapture, MlxHeadForward, MlxRowTable, MlxPromptRange, MlxCacheLayer, MlxSequenceRange, MlxMixedRange, MlxPhaseGroup, MlxMixedGroupRange, MlxSeparateMixed, MlxKnownRange } from "../../contracts/mlx/graph";
|
|
11
|
+
import type { MediaEncoders, MlxPromptInput, Vision, MlxPositionedForward, MlxPositionedRange, MlxPositionedEmbeddingsRange, MlxEmbeddingsRange } from "../../contracts/mlx/media";
|
|
12
|
+
import type { MropeForwardState } from "../../contracts/mlx/positions";
|
|
13
|
+
import type { GraphDeclaration } from "../../contracts/portable/graph";
|
|
14
|
+
import { LoraState } from "../../layers/lora";
|
|
15
|
+
import { RMSNorm } from "../../layers/normalization";
|
|
16
|
+
import { QuantizedEmbedding } from "../../layers/quantized-embedding";
|
|
17
|
+
import { QuantizedLinear, QuantizedLinearPlain as Plain } from "../../layers/quantized-linear";
|
|
18
|
+
import { mropeInvFreq } from "../../layers/qwen-mrope";
|
|
19
|
+
import { ReduceSplitMixedGateUp as MixedGate } from "../../layers/trellis-reduce-gate-up";
|
|
20
|
+
import { CodebookSplitGateUpRow as GateRow } from "../../layers/trellis-codebook-gate-up";
|
|
21
|
+
import { CodebookDownK3iRow as DownK3iRow, CodebookDownRow as DownRow } from "../../layers/trellis-codebook-down";
|
|
22
|
+
import { CodebookMmaGateUp as GateRows, MmaGateUpWide as WideGate } from "../../layers/trellis-gate-up";
|
|
23
|
+
import { CodebookMmaDown as DownRows, CodebookMmaDownK3i as DownK3iRows, MmaDownWide as WideDown, MmaDownWideK3i as WideDownK3i } from "../../layers/trellis-down";
|
|
24
|
+
import { AffineRowLinear as AffineRow, Affine3RowsLinear as Affine3, Affine4RowsLinear as Affine4, AffineSplitMmaLinear as AffineMma } from "../../layers/affine-verify-linear";
|
|
25
|
+
import { DequantizedMatmulLinear as Dense } from "../../layers/dequantized-linear";
|
|
26
|
+
import { AneAffineSplit, AneTrellisExpandedMlpSplit as AneSplit, AneTrellisExpandedMlpSplitK3i as AneSplitK3i, aneAffineSplitProgram, aneMlpSplitProgram } from "../../layers/ane-prefill-split";
|
|
27
|
+
import { TrellisFusedMLP, TrellisUnfusedMLP, TrellisReducePrompt as Reduce, TrellisScatterPrompt as Scatter, TrellisSplitKPrompt as SplitK, TrellisTiledPrompt as Tiled, TrellisExpandReducePrompt as ExpandGate, TrellisExpandScatterPrompt as ExpandDown } from "../../layers/trellis-prompt";
|
|
28
|
+
import { GraphForwardScope, BoundedGraphForwardScope, SubmittedGraphForwardScope, QueuedGraphForwardScope } from "../../runtime/graph-forward";
|
|
29
|
+
import { disposeResources } from "../../runtime/resources";
|
|
30
|
+
import { declareGraph } from "../capabilities";
|
|
31
|
+
import { type ReadPhase } from "../graph";
|
|
32
|
+
import { QwenDeltaLayer as Delta, QwenAttentionLayer as Attention, QwenResidualDeltaLayer as ResidualDelta, QwenResidualAttentionLayer as ResidualAttention, QwenFusedHeadsDeltaNet as FusedDelta } from "./blocks";
|
|
33
|
+
import { bindQwenMediaInput, qwenDraftTarget, qwenDraftVocabularyMemo, qwenMediaEncoders, qwenResidualBasis } from "./media-input";
|
|
34
|
+
import { DeltaWeights, AttentionWeights, GateDownMLP, AneWeightShares, AneBoundMlp, canonicalQwenWeights, type QwenProjectionTarget } from "./written-layers";
|
|
35
|
+
import { PackedMixedScope } from "../../input/packed-mixed";
|
|
36
|
+
|
|
37
|
+
export const QWEN38_TRELLIS_M1MAX_GRAPH="qwen3.8-27b-trellis-3.2bpw-m1max";
|
|
38
|
+
/** Tuned exact entries. Larger rows use the declared operations in this file. */
|
|
39
|
+
export const DECODE_ROWS=256;
|
|
40
|
+
export const VERIFY_ROWS=256;
|
|
41
|
+
export interface AneWidths {hidden:number;intermediate:number;projections:readonly [readonly [number,number],readonly [number,number],readonly [number,number],readonly [number,number]]}
|
|
42
|
+
export const ANE_WIDTHS:AneWidths={hidden:5120,intermediate:17408,projections:[[5120,10240],[5120,6144],[6144,5120],[5120,12288]]};
|
|
43
|
+
/** Programs are installed beforehand. Loading this graph loads these fixed programs. */
|
|
44
|
+
export function anePrograms(widths:AneWidths=ANE_WIDTHS):AneProgramSpec[] {
|
|
45
|
+
const h=widths.hidden,i=widths.intermediate,[[qi,qo],[zi,zo],[oi,oo],[ai,ao]]=widths.projections;
|
|
46
|
+
return [
|
|
47
|
+
aneMlpSplitProgram(h,i,.353,256,4),aneMlpSplitProgram(h,i,.353,320,4),aneMlpSplitProgram(h,i,.353,384,4),
|
|
48
|
+
aneMlpSplitProgram(h,i,.353,448,4),aneMlpSplitProgram(h,i,.353,512,4),aneMlpSplitProgram(h,i,.353,576,4),
|
|
49
|
+
aneMlpSplitProgram(h,i,.353,640,4),aneMlpSplitProgram(h,i,.353,704,4),
|
|
50
|
+
aneMlpSplitProgram(h,i,.353,768,4),
|
|
51
|
+
aneAffineSplitProgram(qi,qo,.375,768,4),aneAffineSplitProgram(zi,zo,.375,768,4),aneAffineSplitProgram(oi,oo,.375,768,4),aneAffineSplitProgram(ai,ao,.375,768,4),
|
|
52
|
+
aneMlpSplitProgram(h,i,.382,896,4),
|
|
53
|
+
aneAffineSplitProgram(qi,qo,.375,896,4),aneAffineSplitProgram(zi,zo,.375,896,4),aneAffineSplitProgram(oi,oo,.375,896,4),aneAffineSplitProgram(ai,ao,.375,896,4),
|
|
54
|
+
aneMlpSplitProgram(h,i,.382,1024,4),
|
|
55
|
+
aneAffineSplitProgram(qi,qo,.375,1024,4),aneAffineSplitProgram(zi,zo,.375,1024,4),aneAffineSplitProgram(oi,oo,.375,1024,4),aneAffineSplitProgram(ai,ao,.375,1024,4),
|
|
56
|
+
aneMlpSplitProgram(h,i,.382,1280,4),
|
|
57
|
+
aneAffineSplitProgram(qi,qo,.375,1280,4),aneAffineSplitProgram(zi,zo,.375,1280,4),aneAffineSplitProgram(oi,oo,.375,1280,4),aneAffineSplitProgram(ai,ao,.375,1280,4),
|
|
58
|
+
aneMlpSplitProgram(h,i,.382,1536,4),
|
|
59
|
+
aneAffineSplitProgram(qi,qo,.375,1536,4),aneAffineSplitProgram(zi,zo,.375,1536,4),aneAffineSplitProgram(oi,oo,.375,1536,4),aneAffineSplitProgram(ai,ao,.375,1536,4),
|
|
60
|
+
aneMlpSplitProgram(h,i,.382,1792,4),
|
|
61
|
+
aneAffineSplitProgram(qi,qo,.375,1792,4),aneAffineSplitProgram(zi,zo,.375,1792,4),aneAffineSplitProgram(oi,oo,.375,1792,4),aneAffineSplitProgram(ai,ao,.375,1792,4),
|
|
62
|
+
aneMlpSplitProgram(h,i,.382,2048,4),
|
|
63
|
+
aneAffineSplitProgram(qi,qo,.375,2048,4),aneAffineSplitProgram(zi,zo,.375,2048,4),aneAffineSplitProgram(oi,oo,.375,2048,4),aneAffineSplitProgram(ai,ao,.375,2048,4),
|
|
64
|
+
];
|
|
65
|
+
}
|
|
66
|
+
const PREFIX="language_model.model";
|
|
67
|
+
const REPRESENTATIONS={gateUpCodebook:true,mixedExactBits:true,scatter:{bits:true,genericBits:true,floatCodebook:true},widePrefill:false};
|
|
68
|
+
type Layers=readonly [
|
|
69
|
+
Delta,Delta,Delta,Attention,Delta,Delta,Delta,Attention,Delta,Delta,Delta,Attention,Delta,Delta,Delta,Attention,
|
|
70
|
+
Delta,Delta,Delta,Attention,Delta,Delta,Delta,Attention,Delta,Delta,Delta,Attention,Delta,Delta,Delta,Attention,
|
|
71
|
+
Delta,Delta,Delta,Attention,Delta,Delta,Delta,Attention,Delta,Delta,Delta,Attention,Delta,Delta,Delta,Attention,
|
|
72
|
+
Delta,Delta,Delta,Attention,Delta,Delta,Delta,Attention,Delta,Delta,Delta,Attention,Delta,Delta,Delta,Attention,
|
|
73
|
+
];
|
|
74
|
+
export type RowsForward=(ids:MlxArray,state:Cache[],capture:MlxLayerCapture|null)=>MlxArray;
|
|
75
|
+
type LayerWeights=readonly [
|
|
76
|
+
DeltaWeights,DeltaWeights,DeltaWeights,AttentionWeights,DeltaWeights,DeltaWeights,DeltaWeights,AttentionWeights,DeltaWeights,DeltaWeights,DeltaWeights,AttentionWeights,DeltaWeights,DeltaWeights,DeltaWeights,AttentionWeights,
|
|
77
|
+
DeltaWeights,DeltaWeights,DeltaWeights,AttentionWeights,DeltaWeights,DeltaWeights,DeltaWeights,AttentionWeights,DeltaWeights,DeltaWeights,DeltaWeights,AttentionWeights,DeltaWeights,DeltaWeights,DeltaWeights,AttentionWeights,
|
|
78
|
+
DeltaWeights,DeltaWeights,DeltaWeights,AttentionWeights,DeltaWeights,DeltaWeights,DeltaWeights,AttentionWeights,DeltaWeights,DeltaWeights,DeltaWeights,AttentionWeights,DeltaWeights,DeltaWeights,DeltaWeights,AttentionWeights,
|
|
79
|
+
DeltaWeights,DeltaWeights,DeltaWeights,AttentionWeights,DeltaWeights,DeltaWeights,DeltaWeights,AttentionWeights,DeltaWeights,DeltaWeights,DeltaWeights,AttentionWeights,DeltaWeights,DeltaWeights,DeltaWeights,AttentionWeights,
|
|
80
|
+
];
|
|
81
|
+
|
|
82
|
+
/** The M1 Max arrangement, with fixed codebook, MMA, and split ANE operations.
|
|
83
|
+
* The loader recommends it; callers may choose it on any machine. */
|
|
84
|
+
export class Qwen38TrellisM1Max implements MlxDeclaredGraph,MlxTokenPhases {
|
|
85
|
+
readonly graph=QWEN38_TRELLIS_M1MAX_GRAPH;
|
|
86
|
+
readonly prefixBase=PREFIX;
|
|
87
|
+
readonly loraState=new LoraState();
|
|
88
|
+
readonly faIdx=3;
|
|
89
|
+
readonly wiredWeights=true;
|
|
90
|
+
readonly weightsBytes:number;
|
|
91
|
+
readonly embed:QuantizedEmbedding;
|
|
92
|
+
readonly finalNorm:RMSNorm;
|
|
93
|
+
readonly lmHead:QuantizedLinear;
|
|
94
|
+
readonly graphCapabilities:GraphDeclaration=declareGraph({media:{input:"embeddings+positions",video:true},hiddenLayerTaps:true,prefill:{boundedWorkspace:true}});
|
|
95
|
+
readonly verifyRoundCosts=new Map([[1,1],[2,1.77],[3,1.95],[4,1.98],[5,2],[6,2.03],[7,2.04],[8,2.05],[16,3.46],[32,6.4]]);
|
|
96
|
+
readonly adaptiveDraftWidths=new Map([["mtp",4]]);
|
|
97
|
+
readonly adaptiveDraftGates=new Map([["dflash2",{decay:.95,idleDecay:.93,probeDepth:1}],["mtp",{decay:.95,idleDecay:.93}]]);
|
|
98
|
+
readonly cacheLayers:readonly MlxCacheLayer[]=[
|
|
99
|
+
{kind:"recurrent"},{kind:"recurrent"},{kind:"recurrent"},{kind:"full",turboQuant:"dense",kv4Head256:true,kv4DecodeBlock:128,windowRead:"folded"},
|
|
100
|
+
{kind:"recurrent"},{kind:"recurrent"},{kind:"recurrent"},{kind:"full",turboQuant:"dense",kv4Head256:true,kv4DecodeBlock:128,windowRead:"folded"},
|
|
101
|
+
{kind:"recurrent"},{kind:"recurrent"},{kind:"recurrent"},{kind:"full",turboQuant:"dense",kv4Head256:true,kv4DecodeBlock:128,windowRead:"folded"},
|
|
102
|
+
{kind:"recurrent"},{kind:"recurrent"},{kind:"recurrent"},{kind:"full",turboQuant:"dense",kv4Head256:true,kv4DecodeBlock:128,windowRead:"folded"},
|
|
103
|
+
{kind:"recurrent"},{kind:"recurrent"},{kind:"recurrent"},{kind:"full",turboQuant:"dense",kv4Head256:true,kv4DecodeBlock:128,windowRead:"folded"},
|
|
104
|
+
{kind:"recurrent"},{kind:"recurrent"},{kind:"recurrent"},{kind:"full",turboQuant:"dense",kv4Head256:true,kv4DecodeBlock:128,windowRead:"folded"},
|
|
105
|
+
{kind:"recurrent"},{kind:"recurrent"},{kind:"recurrent"},{kind:"full",turboQuant:"dense",kv4Head256:true,kv4DecodeBlock:128,windowRead:"folded"},
|
|
106
|
+
{kind:"recurrent"},{kind:"recurrent"},{kind:"recurrent"},{kind:"full",turboQuant:"dense",kv4Head256:true,kv4DecodeBlock:128,windowRead:"folded"},
|
|
107
|
+
{kind:"recurrent"},{kind:"recurrent"},{kind:"recurrent"},{kind:"full",turboQuant:"dense",kv4Head256:true,kv4DecodeBlock:128,windowRead:"folded"},
|
|
108
|
+
{kind:"recurrent"},{kind:"recurrent"},{kind:"recurrent"},{kind:"full",turboQuant:"dense",kv4Head256:true,kv4DecodeBlock:128,windowRead:"folded"},
|
|
109
|
+
{kind:"recurrent"},{kind:"recurrent"},{kind:"recurrent"},{kind:"full",turboQuant:"dense",kv4Head256:true,kv4DecodeBlock:128,windowRead:"folded"},
|
|
110
|
+
{kind:"recurrent"},{kind:"recurrent"},{kind:"recurrent"},{kind:"full",turboQuant:"dense",kv4Head256:true,kv4DecodeBlock:128,windowRead:"folded"},
|
|
111
|
+
{kind:"recurrent"},{kind:"recurrent"},{kind:"recurrent"},{kind:"full",turboQuant:"dense",kv4Head256:true,kv4DecodeBlock:128,windowRead:"folded"},
|
|
112
|
+
{kind:"recurrent"},{kind:"recurrent"},{kind:"recurrent"},{kind:"full",turboQuant:"dense",kv4Head256:true,kv4DecodeBlock:128,windowRead:"folded"},
|
|
113
|
+
{kind:"recurrent"},{kind:"recurrent"},{kind:"recurrent"},{kind:"full",turboQuant:"dense",kv4Head256:true,kv4DecodeBlock:128,windowRead:"folded"},
|
|
114
|
+
{kind:"recurrent"},{kind:"recurrent"},{kind:"recurrent"},{kind:"full",turboQuant:"dense",kv4Head256:true,kv4DecodeBlock:128,windowRead:"folded"},
|
|
115
|
+
];
|
|
116
|
+
readonly prompt:readonly MlxPromptRange[];
|
|
117
|
+
readonly multiSequencePrompt:readonly MlxPromptRange[];
|
|
118
|
+
readonly promptChunk:readonly MlxPromptRange[];
|
|
119
|
+
readonly multiSequencePromptChunk:readonly MlxPromptRange[];
|
|
120
|
+
readonly decode:MlxRowTable<RowsForward>;
|
|
121
|
+
readonly verify:MlxRowTable<RowsForward>;
|
|
122
|
+
readonly appendKnown:readonly MlxKnownRange[];
|
|
123
|
+
readonly logits:MlxRowTable<MlxHeadForward>;
|
|
124
|
+
readonly sequence:readonly MlxSequenceRange[];
|
|
125
|
+
readonly mediaPrefill:readonly MlxPositionedEmbeddingsRange[];
|
|
126
|
+
readonly mediaPrompt:readonly MlxPositionedRange[];
|
|
127
|
+
readonly mediaDecode:MlxRowTable<MlxPositionedForward>;
|
|
128
|
+
readonly embeddingPrompt:readonly MlxEmbeddingsRange[];
|
|
129
|
+
readonly mixed:readonly MlxMixedRange[];
|
|
130
|
+
readonly mixedSeparate:MlxSeparateMixed;
|
|
131
|
+
readonly #head:Plain;
|
|
132
|
+
readonly #targets:Map<string,QwenProjectionTarget>;
|
|
133
|
+
readonly #owned:{dispose():void}[];
|
|
134
|
+
readonly #residualBasis:ResidualBasis|undefined;
|
|
135
|
+
readonly #mrope:{invFreq:MlxArray;rotaryDims:number};
|
|
136
|
+
readonly #one:Layers;
|
|
137
|
+
readonly #rows:Layers;
|
|
138
|
+
readonly #mma:Layers;
|
|
139
|
+
readonly #mmaFive:Layers;
|
|
140
|
+
readonly #wide:Layers;
|
|
141
|
+
readonly #draftProjections:DraftProjectionKernels;
|
|
142
|
+
readonly #vocabulary:()=>NonNullable<DraftProjection["vocabulary"]>;
|
|
143
|
+
readonly #stockNarrow:Layers;
|
|
144
|
+
readonly #fiveToEight:Layers;
|
|
145
|
+
readonly #tiled:Layers;
|
|
146
|
+
readonly #expanded:Layers;
|
|
147
|
+
|
|
148
|
+
constructor(weights:Weights,readonly config:ModelConfig) {
|
|
149
|
+
const w=canonicalQwenWeights(weights),c=config,t=config.text;
|
|
150
|
+
this.weightsBytes=[...weights.shards.files.values()].reduce((bytes,file)=>bytes+file.mmap.size,0);
|
|
151
|
+
this.#residualBasis=qwenResidualBasis(this.graph,config.modelDir);
|
|
152
|
+
this.embed=QuantizedEmbedding.load(w,`${PREFIX}.embed_tokens`,c);
|
|
153
|
+
this.finalNorm=new RMSNorm(w.tensor(`${PREFIX}.norm.weight`),t.rmsNormEps);
|
|
154
|
+
this.lmHead=QuantizedLinear.load(w,"language_model.lm_head",c);
|
|
155
|
+
this.#head=new Plain(this.lmHead);
|
|
156
|
+
this.#draftProjections=[
|
|
157
|
+
{from:1,to:1,operation:(w,scales,biases)=>{const kernel=new Plain(new QuantizedLinear(w,scales,biases,{bits:4,groupSize:64,mode:"affine"}));return x=>kernel.forward(x);}},
|
|
158
|
+
{from:2,to:3,operation:(w,scales,biases)=>{const kernel=new Affine4(new QuantizedLinear(w,scales,biases,{bits:4,groupSize:64,mode:"affine"}));return x=>kernel.forward(x);}},
|
|
159
|
+
{from:4,to:Infinity,operation:(w,scales,biases)=>{const kernel=new AffineMma(new QuantizedLinear(w,scales,biases,{bits:4,groupSize:64,mode:"affine"}));return x=>kernel.forward(x);}},
|
|
160
|
+
];
|
|
161
|
+
const invScale=Math.pow(t.linearKeyHeadDim,-0.5),qk={q:ops.filledBf16(invScale*invScale,t.linearKeyHeadDim),k:ops.filledBf16(invScale,t.linearKeyHeadDim)};
|
|
162
|
+
const rotaryDims=Math.trunc(t.headDim*t.partialRotaryFactor);
|
|
163
|
+
this.#mrope={invFreq:mropeInvFreq(rotaryDims,t.ropeParameters.full_attention!.ropeTheta),rotaryDims};
|
|
164
|
+
this.#owned=[qk.q,qk.k,this.#mrope.invFreq];
|
|
165
|
+
this.#vocabulary=qwenDraftVocabularyMemo(this.lmHead,c.modelDir,"mtp/draft_vocab.json",t.vocabSize,head=>{
|
|
166
|
+
const row=new AffineRow(head),rows=new Affine4(head),mma=new AffineMma(head),plain=new Plain(head);
|
|
167
|
+
return [{from:1,to:1,operation:hidden=>row.forward(hidden)},{from:2,to:3,operation:hidden=>rows.forward(hidden)},
|
|
168
|
+
{from:4,to:256,operation:hidden=>mma.forward(hidden)},{from:257,to:Infinity,operation:hidden=>plain.forward(hidden)}];
|
|
169
|
+
},this.#owned);
|
|
170
|
+
const state=this.loraState;
|
|
171
|
+
const d0=new DeltaWeights(w,c,`${PREFIX}.layers.0`,REPRESENTATIONS,state);
|
|
172
|
+
const d1=new DeltaWeights(w,c,`${PREFIX}.layers.1`,REPRESENTATIONS,state);
|
|
173
|
+
const d2=new DeltaWeights(w,c,`${PREFIX}.layers.2`,REPRESENTATIONS,state);
|
|
174
|
+
const a3=new AttentionWeights(w,c,`${PREFIX}.layers.3`,REPRESENTATIONS,state);
|
|
175
|
+
const d4=new DeltaWeights(w,c,`${PREFIX}.layers.4`,REPRESENTATIONS,state);
|
|
176
|
+
const d5=new DeltaWeights(w,c,`${PREFIX}.layers.5`,REPRESENTATIONS,state);
|
|
177
|
+
const d6=new DeltaWeights(w,c,`${PREFIX}.layers.6`,REPRESENTATIONS,state);
|
|
178
|
+
const a7=new AttentionWeights(w,c,`${PREFIX}.layers.7`,REPRESENTATIONS,state);
|
|
179
|
+
const d8=new DeltaWeights(w,c,`${PREFIX}.layers.8`,REPRESENTATIONS,state);
|
|
180
|
+
const d9=new DeltaWeights(w,c,`${PREFIX}.layers.9`,REPRESENTATIONS,state);
|
|
181
|
+
const d10=new DeltaWeights(w,c,`${PREFIX}.layers.10`,REPRESENTATIONS,state);
|
|
182
|
+
const a11=new AttentionWeights(w,c,`${PREFIX}.layers.11`,REPRESENTATIONS,state);
|
|
183
|
+
const d12=new DeltaWeights(w,c,`${PREFIX}.layers.12`,REPRESENTATIONS,state);
|
|
184
|
+
const d13=new DeltaWeights(w,c,`${PREFIX}.layers.13`,REPRESENTATIONS,state);
|
|
185
|
+
const d14=new DeltaWeights(w,c,`${PREFIX}.layers.14`,REPRESENTATIONS,state);
|
|
186
|
+
const a15=new AttentionWeights(w,c,`${PREFIX}.layers.15`,REPRESENTATIONS,state);
|
|
187
|
+
const d16=new DeltaWeights(w,c,`${PREFIX}.layers.16`,REPRESENTATIONS,state);
|
|
188
|
+
const d17=new DeltaWeights(w,c,`${PREFIX}.layers.17`,REPRESENTATIONS,state);
|
|
189
|
+
const d18=new DeltaWeights(w,c,`${PREFIX}.layers.18`,REPRESENTATIONS,state);
|
|
190
|
+
const a19=new AttentionWeights(w,c,`${PREFIX}.layers.19`,REPRESENTATIONS,state);
|
|
191
|
+
const d20=new DeltaWeights(w,c,`${PREFIX}.layers.20`,REPRESENTATIONS,state);
|
|
192
|
+
const d21=new DeltaWeights(w,c,`${PREFIX}.layers.21`,REPRESENTATIONS,state);
|
|
193
|
+
const d22=new DeltaWeights(w,c,`${PREFIX}.layers.22`,REPRESENTATIONS,state);
|
|
194
|
+
const a23=new AttentionWeights(w,c,`${PREFIX}.layers.23`,REPRESENTATIONS,state);
|
|
195
|
+
const d24=new DeltaWeights(w,c,`${PREFIX}.layers.24`,REPRESENTATIONS,state);
|
|
196
|
+
const d25=new DeltaWeights(w,c,`${PREFIX}.layers.25`,REPRESENTATIONS,state);
|
|
197
|
+
const d26=new DeltaWeights(w,c,`${PREFIX}.layers.26`,REPRESENTATIONS,state);
|
|
198
|
+
const a27=new AttentionWeights(w,c,`${PREFIX}.layers.27`,REPRESENTATIONS,state);
|
|
199
|
+
const d28=new DeltaWeights(w,c,`${PREFIX}.layers.28`,REPRESENTATIONS,state);
|
|
200
|
+
const d29=new DeltaWeights(w,c,`${PREFIX}.layers.29`,REPRESENTATIONS,state);
|
|
201
|
+
const d30=new DeltaWeights(w,c,`${PREFIX}.layers.30`,REPRESENTATIONS,state);
|
|
202
|
+
const a31=new AttentionWeights(w,c,`${PREFIX}.layers.31`,REPRESENTATIONS,state);
|
|
203
|
+
const d32=new DeltaWeights(w,c,`${PREFIX}.layers.32`,REPRESENTATIONS,state);
|
|
204
|
+
const d33=new DeltaWeights(w,c,`${PREFIX}.layers.33`,REPRESENTATIONS,state);
|
|
205
|
+
const d34=new DeltaWeights(w,c,`${PREFIX}.layers.34`,REPRESENTATIONS,state);
|
|
206
|
+
const a35=new AttentionWeights(w,c,`${PREFIX}.layers.35`,REPRESENTATIONS,state);
|
|
207
|
+
const d36=new DeltaWeights(w,c,`${PREFIX}.layers.36`,REPRESENTATIONS,state);
|
|
208
|
+
const d37=new DeltaWeights(w,c,`${PREFIX}.layers.37`,REPRESENTATIONS,state);
|
|
209
|
+
const d38=new DeltaWeights(w,c,`${PREFIX}.layers.38`,REPRESENTATIONS,state);
|
|
210
|
+
const a39=new AttentionWeights(w,c,`${PREFIX}.layers.39`,REPRESENTATIONS,state);
|
|
211
|
+
const d40=new DeltaWeights(w,c,`${PREFIX}.layers.40`,REPRESENTATIONS,state);
|
|
212
|
+
const d41=new DeltaWeights(w,c,`${PREFIX}.layers.41`,REPRESENTATIONS,state);
|
|
213
|
+
const d42=new DeltaWeights(w,c,`${PREFIX}.layers.42`,REPRESENTATIONS,state);
|
|
214
|
+
const a43=new AttentionWeights(w,c,`${PREFIX}.layers.43`,REPRESENTATIONS,state);
|
|
215
|
+
const d44=new DeltaWeights(w,c,`${PREFIX}.layers.44`,REPRESENTATIONS,state);
|
|
216
|
+
const d45=new DeltaWeights(w,c,`${PREFIX}.layers.45`,REPRESENTATIONS,state);
|
|
217
|
+
const d46=new DeltaWeights(w,c,`${PREFIX}.layers.46`,REPRESENTATIONS,state);
|
|
218
|
+
const a47=new AttentionWeights(w,c,`${PREFIX}.layers.47`,REPRESENTATIONS,state);
|
|
219
|
+
const d48=new DeltaWeights(w,c,`${PREFIX}.layers.48`,REPRESENTATIONS,state);
|
|
220
|
+
const d49=new DeltaWeights(w,c,`${PREFIX}.layers.49`,REPRESENTATIONS,state);
|
|
221
|
+
const d50=new DeltaWeights(w,c,`${PREFIX}.layers.50`,REPRESENTATIONS,state);
|
|
222
|
+
const a51=new AttentionWeights(w,c,`${PREFIX}.layers.51`,REPRESENTATIONS,state);
|
|
223
|
+
const d52=new DeltaWeights(w,c,`${PREFIX}.layers.52`,REPRESENTATIONS,state);
|
|
224
|
+
const d53=new DeltaWeights(w,c,`${PREFIX}.layers.53`,REPRESENTATIONS,state);
|
|
225
|
+
const d54=new DeltaWeights(w,c,`${PREFIX}.layers.54`,REPRESENTATIONS,state);
|
|
226
|
+
const a55=new AttentionWeights(w,c,`${PREFIX}.layers.55`,REPRESENTATIONS,state);
|
|
227
|
+
const d56=new DeltaWeights(w,c,`${PREFIX}.layers.56`,REPRESENTATIONS,state);
|
|
228
|
+
const d57=new DeltaWeights(w,c,`${PREFIX}.layers.57`,REPRESENTATIONS,state);
|
|
229
|
+
const d58=new DeltaWeights(w,c,`${PREFIX}.layers.58`,REPRESENTATIONS,state);
|
|
230
|
+
const a59=new AttentionWeights(w,c,`${PREFIX}.layers.59`,REPRESENTATIONS,state);
|
|
231
|
+
const d60=new DeltaWeights(w,c,`${PREFIX}.layers.60`,REPRESENTATIONS,state);
|
|
232
|
+
const d61=new DeltaWeights(w,c,`${PREFIX}.layers.61`,REPRESENTATIONS,state);
|
|
233
|
+
const d62=new DeltaWeights(w,c,`${PREFIX}.layers.62`,REPRESENTATIONS,state);
|
|
234
|
+
const a63=new AttentionWeights(w,c,`${PREFIX}.layers.63`,REPRESENTATIONS,state);
|
|
235
|
+
this.#owned.push(d0.trellis,d1.trellis,d2.trellis,a3.trellis,d4.trellis,d5.trellis,d6.trellis,a7.trellis,
|
|
236
|
+
d8.trellis,d9.trellis,d10.trellis,a11.trellis,d12.trellis,d13.trellis,d14.trellis,a15.trellis,
|
|
237
|
+
d16.trellis,d17.trellis,d18.trellis,a19.trellis,d20.trellis,d21.trellis,d22.trellis,a23.trellis,
|
|
238
|
+
d24.trellis,d25.trellis,d26.trellis,a27.trellis,d28.trellis,d29.trellis,d30.trellis,a31.trellis,
|
|
239
|
+
d32.trellis,d33.trellis,d34.trellis,a35.trellis,d36.trellis,d37.trellis,d38.trellis,a39.trellis,
|
|
240
|
+
d40.trellis,d41.trellis,d42.trellis,a43.trellis,d44.trellis,d45.trellis,d46.trellis,a47.trellis,
|
|
241
|
+
d48.trellis,d49.trellis,d50.trellis,a51.trellis,d52.trellis,d53.trellis,d54.trellis,a55.trellis,
|
|
242
|
+
d56.trellis,d57.trellis,d58.trellis,a59.trellis,d60.trellis,d61.trellis,d62.trellis,a63.trellis);
|
|
243
|
+
this.#targets=new Map([...d0.targets(),...d1.targets(),...d2.targets(),...a3.targets(),...d4.targets(),...d5.targets(),...d6.targets(),...a7.targets(),
|
|
244
|
+
...d8.targets(),...d9.targets(),...d10.targets(),...a11.targets(),...d12.targets(),...d13.targets(),...d14.targets(),...a15.targets(),
|
|
245
|
+
...d16.targets(),...d17.targets(),...d18.targets(),...a19.targets(),...d20.targets(),...d21.targets(),...d22.targets(),...a23.targets(),
|
|
246
|
+
...d24.targets(),...d25.targets(),...d26.targets(),...a27.targets(),...d28.targets(),...d29.targets(),...d30.targets(),...a31.targets(),
|
|
247
|
+
...d32.targets(),...d33.targets(),...d34.targets(),...a35.targets(),...d36.targets(),...d37.targets(),...d38.targets(),...a39.targets(),
|
|
248
|
+
...d40.targets(),...d41.targets(),...d42.targets(),...a43.targets(),...d44.targets(),...d45.targets(),...d46.targets(),...a47.targets(),
|
|
249
|
+
...d48.targets(),...d49.targets(),...d50.targets(),...a51.targets(),...d52.targets(),...d53.targets(),...d54.targets(),...a55.targets(),
|
|
250
|
+
...d56.targets(),...d57.targets(),...d58.targets(),...a59.targets(),...d60.targets(),...d61.targets(),...d62.targets(),...a63.targets()]);
|
|
251
|
+
this.#one=[
|
|
252
|
+
d0.stage(w,c,ResidualDelta,GateDownMLP,GateRow,Reduce,DownRow,Plain,Plain,Plain,Plain,Plain,qk,FusedDelta),
|
|
253
|
+
d1.stage(w,c,ResidualDelta,GateDownMLP,GateRow,Reduce,DownK3iRow,AffineRow,AffineRow,Plain,Plain,Plain,qk,FusedDelta),
|
|
254
|
+
d2.stage(w,c,ResidualDelta,GateDownMLP,GateRow,Reduce,DownK3iRow,AffineRow,AffineRow,Plain,Plain,Plain,qk,FusedDelta),
|
|
255
|
+
a3.stage(w,c,ResidualAttention,GateDownMLP,GateRow,Reduce,DownK3iRow,Plain,Plain,Plain,Plain),
|
|
256
|
+
d4.stage(w,c,ResidualDelta,GateDownMLP,GateRow,Reduce,DownK3iRow,AffineRow,Plain,Plain,Plain,Plain,qk,FusedDelta),
|
|
257
|
+
d5.stage(w,c,ResidualDelta,GateDownMLP,GateRow,Reduce,DownK3iRow,AffineRow,Plain,Plain,Plain,Plain,qk,FusedDelta),
|
|
258
|
+
d6.stage(w,c,ResidualDelta,GateDownMLP,GateRow,Reduce,DownK3iRow,AffineRow,AffineRow,Plain,Plain,AffineRow,qk,FusedDelta),
|
|
259
|
+
a7.stage(w,c,ResidualAttention,GateDownMLP,GateRow,Reduce,DownK3iRow,Plain,Plain,Plain,Plain),
|
|
260
|
+
d8.stage(w,c,ResidualDelta,GateDownMLP,GateRow,Reduce,DownK3iRow,AffineRow,AffineRow,Plain,Plain,AffineRow,qk,FusedDelta),
|
|
261
|
+
d9.stage(w,c,ResidualDelta,GateDownMLP,GateRow,Reduce,DownK3iRow,AffineRow,AffineRow,Plain,Plain,AffineRow,qk,FusedDelta),
|
|
262
|
+
d10.stage(w,c,ResidualDelta,GateDownMLP,GateRow,Reduce,DownK3iRow,AffineRow,AffineRow,Plain,Plain,AffineRow,qk,FusedDelta),
|
|
263
|
+
a11.stage(w,c,ResidualAttention,GateDownMLP,GateRow,Reduce,DownRow,Plain,Plain,Plain,Plain),
|
|
264
|
+
d12.stage(w,c,ResidualDelta,GateDownMLP,GateRow,Reduce,DownK3iRow,AffineRow,AffineRow,Plain,Plain,AffineRow,qk,FusedDelta),
|
|
265
|
+
d13.stage(w,c,ResidualDelta,GateDownMLP,GateRow,Reduce,DownRow,AffineRow,AffineRow,Plain,Plain,AffineRow,qk,FusedDelta),
|
|
266
|
+
d14.stage(w,c,ResidualDelta,GateDownMLP,GateRow,Reduce,DownK3iRow,AffineRow,AffineRow,Plain,Plain,AffineRow,qk,FusedDelta),
|
|
267
|
+
a15.stage(w,c,ResidualAttention,GateDownMLP,GateRow,Reduce,DownRow,Plain,Plain,Plain,Plain),
|
|
268
|
+
d16.stage(w,c,ResidualDelta,GateDownMLP,GateRow,Reduce,DownRow,AffineRow,AffineRow,Plain,Plain,AffineRow,qk,FusedDelta),
|
|
269
|
+
d17.stage(w,c,ResidualDelta,GateDownMLP,GateRow,Reduce,DownRow,AffineRow,AffineRow,Plain,Plain,AffineRow,qk,FusedDelta),
|
|
270
|
+
d18.stage(w,c,ResidualDelta,GateDownMLP,GateRow,Reduce,DownRow,AffineRow,AffineRow,Plain,Plain,AffineRow,qk,FusedDelta),
|
|
271
|
+
a19.stage(w,c,ResidualAttention,GateDownMLP,GateRow,Reduce,DownK3iRow,Plain,Plain,Plain,Plain),
|
|
272
|
+
d20.stage(w,c,ResidualDelta,GateDownMLP,GateRow,Reduce,DownK3iRow,AffineRow,AffineRow,Plain,Plain,AffineRow,qk,FusedDelta),
|
|
273
|
+
d21.stage(w,c,ResidualDelta,GateDownMLP,GateRow,Reduce,DownK3iRow,AffineRow,AffineRow,Plain,Plain,AffineRow,qk,FusedDelta),
|
|
274
|
+
d22.stage(w,c,ResidualDelta,GateDownMLP,GateRow,Reduce,DownK3iRow,AffineRow,AffineRow,Plain,Plain,AffineRow,qk,FusedDelta),
|
|
275
|
+
a23.stage(w,c,ResidualAttention,GateDownMLP,GateRow,Reduce,DownK3iRow,Plain,Plain,Plain,Plain),
|
|
276
|
+
d24.stage(w,c,ResidualDelta,GateDownMLP,GateRow,Reduce,DownK3iRow,AffineRow,AffineRow,Plain,Plain,AffineRow,qk,FusedDelta),
|
|
277
|
+
d25.stage(w,c,ResidualDelta,GateDownMLP,GateRow,Reduce,DownK3iRow,AffineRow,Plain,Plain,Plain,AffineRow,qk,FusedDelta),
|
|
278
|
+
d26.stage(w,c,ResidualDelta,GateDownMLP,GateRow,Reduce,DownK3iRow,AffineRow,AffineRow,Plain,Plain,AffineRow,qk,FusedDelta),
|
|
279
|
+
a27.stage(w,c,ResidualAttention,GateDownMLP,GateRow,Reduce,DownK3iRow,Plain,Plain,Plain,Plain),
|
|
280
|
+
d28.stage(w,c,ResidualDelta,GateDownMLP,GateRow,Reduce,DownK3iRow,AffineRow,AffineRow,Plain,Plain,AffineRow,qk,FusedDelta),
|
|
281
|
+
d29.stage(w,c,ResidualDelta,GateDownMLP,GateRow,Reduce,DownRow,AffineRow,AffineRow,Plain,Plain,AffineRow,qk,FusedDelta),
|
|
282
|
+
d30.stage(w,c,ResidualDelta,GateDownMLP,GateRow,Reduce,DownK3iRow,AffineRow,AffineRow,Plain,Plain,AffineRow,qk,FusedDelta),
|
|
283
|
+
a31.stage(w,c,ResidualAttention,GateDownMLP,GateRow,Reduce,DownRow,Plain,Plain,Plain,Plain),
|
|
284
|
+
d32.stage(w,c,ResidualDelta,GateDownMLP,GateRow,Reduce,DownRow,AffineRow,AffineRow,Plain,Plain,AffineRow,qk,FusedDelta),
|
|
285
|
+
d33.stage(w,c,ResidualDelta,GateDownMLP,GateRow,Reduce,DownRow,AffineRow,AffineRow,Plain,Plain,AffineRow,qk,FusedDelta),
|
|
286
|
+
d34.stage(w,c,ResidualDelta,GateDownMLP,GateRow,Reduce,DownK3iRow,AffineRow,AffineRow,Plain,Plain,AffineRow,qk,FusedDelta),
|
|
287
|
+
a35.stage(w,c,ResidualAttention,GateDownMLP,GateRow,Reduce,DownK3iRow,Plain,Plain,Plain,Plain),
|
|
288
|
+
d36.stage(w,c,ResidualDelta,GateDownMLP,GateRow,Reduce,DownK3iRow,AffineRow,AffineRow,Plain,Plain,AffineRow,qk,FusedDelta),
|
|
289
|
+
d37.stage(w,c,ResidualDelta,GateDownMLP,GateRow,Reduce,DownK3iRow,AffineRow,AffineRow,Plain,Plain,AffineRow,qk,FusedDelta),
|
|
290
|
+
d38.stage(w,c,ResidualDelta,GateDownMLP,GateRow,Reduce,DownK3iRow,AffineRow,AffineRow,Plain,Plain,AffineRow,qk,FusedDelta),
|
|
291
|
+
a39.stage(w,c,ResidualAttention,GateDownMLP,GateRow,Reduce,DownK3iRow,Plain,Plain,Plain,Plain),
|
|
292
|
+
d40.stage(w,c,ResidualDelta,GateDownMLP,GateRow,Reduce,DownK3iRow,AffineRow,AffineRow,Plain,Plain,AffineRow,qk,FusedDelta),
|
|
293
|
+
d41.stage(w,c,ResidualDelta,GateDownMLP,GateRow,Reduce,DownK3iRow,AffineRow,AffineRow,Plain,Plain,AffineRow,qk,FusedDelta),
|
|
294
|
+
d42.stage(w,c,ResidualDelta,GateDownMLP,GateRow,Reduce,DownK3iRow,AffineRow,AffineRow,Plain,Plain,AffineRow,qk,FusedDelta),
|
|
295
|
+
a43.stage(w,c,ResidualAttention,GateDownMLP,GateRow,Reduce,DownK3iRow,Plain,Plain,Plain,Plain),
|
|
296
|
+
d44.stage(w,c,ResidualDelta,GateDownMLP,GateRow,Reduce,DownK3iRow,AffineRow,AffineRow,Plain,Plain,AffineRow,qk,FusedDelta),
|
|
297
|
+
d45.stage(w,c,ResidualDelta,GateDownMLP,GateRow,Reduce,DownRow,AffineRow,AffineRow,Plain,Plain,AffineRow,qk,FusedDelta),
|
|
298
|
+
d46.stage(w,c,ResidualDelta,GateDownMLP,GateRow,Reduce,DownRow,AffineRow,AffineRow,Plain,Plain,AffineRow,qk,FusedDelta),
|
|
299
|
+
a47.stage(w,c,ResidualAttention,GateDownMLP,GateRow,Reduce,DownRow,Plain,Plain,Plain,Plain),
|
|
300
|
+
d48.stage(w,c,ResidualDelta,GateDownMLP,GateRow,Reduce,DownRow,AffineRow,AffineRow,Plain,Plain,AffineRow,qk,FusedDelta),
|
|
301
|
+
d49.stage(w,c,ResidualDelta,GateDownMLP,GateRow,Reduce,DownRow,AffineRow,AffineRow,Plain,Plain,AffineRow,qk,FusedDelta),
|
|
302
|
+
d50.stage(w,c,ResidualDelta,GateDownMLP,GateRow,Reduce,DownK3iRow,AffineRow,AffineRow,Plain,Plain,AffineRow,qk,FusedDelta),
|
|
303
|
+
a51.stage(w,c,ResidualAttention,GateDownMLP,GateRow,Reduce,DownK3iRow,Plain,Plain,Plain,Plain),
|
|
304
|
+
d52.stage(w,c,ResidualDelta,GateDownMLP,GateRow,Reduce,DownK3iRow,AffineRow,AffineRow,Plain,Plain,AffineRow,qk,FusedDelta),
|
|
305
|
+
d53.stage(w,c,ResidualDelta,GateDownMLP,GateRow,Reduce,DownK3iRow,AffineRow,Plain,Plain,Plain,AffineRow,qk,FusedDelta),
|
|
306
|
+
d54.stage(w,c,ResidualDelta,GateDownMLP,GateRow,Reduce,DownRow,Plain,Plain,Plain,Plain,Plain,qk,FusedDelta),
|
|
307
|
+
a55.stage(w,c,ResidualAttention,GateDownMLP,GateRow,Reduce,DownRow,Plain,Plain,Plain,Plain),
|
|
308
|
+
d56.stage(w,c,ResidualDelta,GateDownMLP,GateRow,Reduce,DownRow,Plain,Plain,Plain,Plain,Plain,qk,FusedDelta),
|
|
309
|
+
d57.stage(w,c,ResidualDelta,GateDownMLP,GateRow,Reduce,DownRow,Plain,Plain,Plain,Plain,Plain,qk,FusedDelta),
|
|
310
|
+
d58.stage(w,c,ResidualDelta,GateDownMLP,GateRow,Reduce,DownRow,Plain,AffineRow,Plain,Plain,AffineRow,qk,FusedDelta),
|
|
311
|
+
a59.stage(w,c,ResidualAttention,GateDownMLP,GateRow,Reduce,DownRow,Plain,Plain,Plain,Plain),
|
|
312
|
+
d60.stage(w,c,ResidualDelta,GateDownMLP,GateRow,Reduce,DownRow,Plain,Plain,Plain,Plain,AffineRow,qk,FusedDelta),
|
|
313
|
+
d61.stage(w,c,ResidualDelta,GateDownMLP,GateRow,Reduce,DownRow,Plain,Plain,Plain,Plain,Plain,qk,FusedDelta),
|
|
314
|
+
d62.stage(w,c,ResidualDelta,GateDownMLP,GateRow,Reduce,DownK3iRow,Plain,Plain,Plain,Plain,AffineRow,qk,FusedDelta),
|
|
315
|
+
a63.stage(w,c,ResidualAttention,GateDownMLP,GateRow,Reduce,DownK3iRow,Plain,Plain,Plain,Plain),
|
|
316
|
+
];
|
|
317
|
+
this.#rows=[
|
|
318
|
+
d0.stage(w,c,Delta,GateDownMLP,GateRows,Reduce,DownRows,Affine4,Affine4,Affine4,Affine4,Affine4,qk),
|
|
319
|
+
d1.stage(w,c,Delta,GateDownMLP,GateRows,Reduce,DownK3iRows,Affine3,Affine3,Affine4,Affine4,Affine4,qk),
|
|
320
|
+
d2.stage(w,c,Delta,GateDownMLP,GateRows,Reduce,DownK3iRows,Affine3,Affine3,Affine4,Affine4,Affine4,qk),
|
|
321
|
+
a3.stage(w,c,Attention,GateDownMLP,GateRows,Reduce,DownK3iRows,Affine4,Affine4,Affine4,Affine4),
|
|
322
|
+
d4.stage(w,c,Delta,GateDownMLP,GateRows,Reduce,DownK3iRows,Affine3,Affine4,Affine4,Affine4,Affine4,qk),
|
|
323
|
+
d5.stage(w,c,Delta,GateDownMLP,GateRows,Reduce,DownK3iRows,Affine3,Affine4,Affine4,Affine4,Affine4,qk),
|
|
324
|
+
d6.stage(w,c,Delta,GateDownMLP,GateRows,Reduce,DownK3iRows,Affine3,Affine3,Affine4,Affine4,Affine3,qk),
|
|
325
|
+
a7.stage(w,c,Attention,GateDownMLP,GateRows,Reduce,DownK3iRows,Affine4,Affine4,Affine4,Affine4),
|
|
326
|
+
d8.stage(w,c,Delta,GateDownMLP,GateRows,Reduce,DownK3iRows,Affine3,Affine3,Affine4,Affine4,Affine3,qk),
|
|
327
|
+
d9.stage(w,c,Delta,GateDownMLP,GateRows,Reduce,DownK3iRows,Affine3,Affine3,Affine4,Affine4,Affine3,qk),
|
|
328
|
+
d10.stage(w,c,Delta,GateDownMLP,GateRows,Reduce,DownK3iRows,Affine3,Affine3,Affine4,Affine4,Affine3,qk),
|
|
329
|
+
a11.stage(w,c,Attention,GateDownMLP,GateRows,Reduce,DownRows,Affine4,Affine4,Affine4,Affine4),
|
|
330
|
+
d12.stage(w,c,Delta,GateDownMLP,GateRows,Reduce,DownK3iRows,Affine3,Affine3,Affine4,Affine4,Affine3,qk),
|
|
331
|
+
d13.stage(w,c,Delta,GateDownMLP,GateRows,Reduce,DownRows,Affine3,Affine3,Affine4,Affine4,Affine3,qk),
|
|
332
|
+
d14.stage(w,c,Delta,GateDownMLP,GateRows,Reduce,DownK3iRows,Affine3,Affine3,Affine4,Affine4,Affine3,qk),
|
|
333
|
+
a15.stage(w,c,Attention,GateDownMLP,GateRows,Reduce,DownRows,Affine4,Affine4,Affine4,Affine4),
|
|
334
|
+
d16.stage(w,c,Delta,GateDownMLP,GateRows,Reduce,DownRows,Affine3,Affine3,Affine4,Affine4,Affine3,qk),
|
|
335
|
+
d17.stage(w,c,Delta,GateDownMLP,GateRows,Reduce,DownRows,Affine3,Affine3,Affine4,Affine4,Affine3,qk),
|
|
336
|
+
d18.stage(w,c,Delta,GateDownMLP,GateRows,Reduce,DownRows,Affine3,Affine3,Affine4,Affine4,Affine3,qk),
|
|
337
|
+
a19.stage(w,c,Attention,GateDownMLP,GateRows,Reduce,DownK3iRows,Affine4,Affine4,Affine4,Affine4),
|
|
338
|
+
d20.stage(w,c,Delta,GateDownMLP,GateRows,Reduce,DownK3iRows,Affine3,Affine3,Affine4,Affine4,Affine3,qk),
|
|
339
|
+
d21.stage(w,c,Delta,GateDownMLP,GateRows,Reduce,DownK3iRows,Affine3,Affine3,Affine4,Affine4,Affine3,qk),
|
|
340
|
+
d22.stage(w,c,Delta,GateDownMLP,GateRows,Reduce,DownK3iRows,Affine3,Affine3,Affine4,Affine4,Affine3,qk),
|
|
341
|
+
a23.stage(w,c,Attention,GateDownMLP,GateRows,Reduce,DownK3iRows,Affine4,Affine4,Affine4,Affine4),
|
|
342
|
+
d24.stage(w,c,Delta,GateDownMLP,GateRows,Reduce,DownK3iRows,Affine3,Affine3,Affine4,Affine4,Affine3,qk),
|
|
343
|
+
d25.stage(w,c,Delta,GateDownMLP,GateRows,Reduce,DownK3iRows,Affine3,Affine4,Affine4,Affine4,Affine3,qk),
|
|
344
|
+
d26.stage(w,c,Delta,GateDownMLP,GateRows,Reduce,DownK3iRows,Affine3,Affine3,Affine4,Affine4,Affine3,qk),
|
|
345
|
+
a27.stage(w,c,Attention,GateDownMLP,GateRows,Reduce,DownK3iRows,Affine4,Affine4,Affine4,Affine4),
|
|
346
|
+
d28.stage(w,c,Delta,GateDownMLP,GateRows,Reduce,DownK3iRows,Affine3,Affine3,Affine4,Affine4,Affine3,qk),
|
|
347
|
+
d29.stage(w,c,Delta,GateDownMLP,GateRows,Reduce,DownRows,Affine3,Affine3,Affine4,Affine4,Affine3,qk),
|
|
348
|
+
d30.stage(w,c,Delta,GateDownMLP,GateRows,Reduce,DownK3iRows,Affine3,Affine3,Affine4,Affine4,Affine3,qk),
|
|
349
|
+
a31.stage(w,c,Attention,GateDownMLP,GateRows,Reduce,DownRows,Affine4,Affine4,Affine4,Affine4),
|
|
350
|
+
d32.stage(w,c,Delta,GateDownMLP,GateRows,Reduce,DownRows,Affine3,Affine3,Affine4,Affine4,Affine3,qk),
|
|
351
|
+
d33.stage(w,c,Delta,GateDownMLP,GateRows,Reduce,DownRows,Affine3,Affine3,Affine4,Affine4,Affine3,qk),
|
|
352
|
+
d34.stage(w,c,Delta,GateDownMLP,GateRows,Reduce,DownK3iRows,Affine3,Affine3,Affine4,Affine4,Affine3,qk),
|
|
353
|
+
a35.stage(w,c,Attention,GateDownMLP,GateRows,Reduce,DownK3iRows,Affine4,Affine4,Affine4,Affine4),
|
|
354
|
+
d36.stage(w,c,Delta,GateDownMLP,GateRows,Reduce,DownK3iRows,Affine3,Affine3,Affine4,Affine4,Affine3,qk),
|
|
355
|
+
d37.stage(w,c,Delta,GateDownMLP,GateRows,Reduce,DownK3iRows,Affine3,Affine3,Affine4,Affine4,Affine3,qk),
|
|
356
|
+
d38.stage(w,c,Delta,GateDownMLP,GateRows,Reduce,DownK3iRows,Affine3,Affine3,Affine4,Affine4,Affine3,qk),
|
|
357
|
+
a39.stage(w,c,Attention,GateDownMLP,GateRows,Reduce,DownK3iRows,Affine4,Affine4,Affine4,Affine4),
|
|
358
|
+
d40.stage(w,c,Delta,GateDownMLP,GateRows,Reduce,DownK3iRows,Affine3,Affine3,Affine4,Affine4,Affine3,qk),
|
|
359
|
+
d41.stage(w,c,Delta,GateDownMLP,GateRows,Reduce,DownK3iRows,Affine3,Affine3,Affine4,Affine4,Affine3,qk),
|
|
360
|
+
d42.stage(w,c,Delta,GateDownMLP,GateRows,Reduce,DownK3iRows,Affine3,Affine3,Affine4,Affine4,Affine3,qk),
|
|
361
|
+
a43.stage(w,c,Attention,GateDownMLP,GateRows,Reduce,DownK3iRows,Affine4,Affine4,Affine4,Affine4),
|
|
362
|
+
d44.stage(w,c,Delta,GateDownMLP,GateRows,Reduce,DownK3iRows,Affine3,Affine3,Affine4,Affine4,Affine3,qk),
|
|
363
|
+
d45.stage(w,c,Delta,GateDownMLP,GateRows,Reduce,DownRows,Affine3,Affine3,Affine4,Affine4,Affine3,qk),
|
|
364
|
+
d46.stage(w,c,Delta,GateDownMLP,GateRows,Reduce,DownRows,Affine3,Affine3,Affine4,Affine4,Affine3,qk),
|
|
365
|
+
a47.stage(w,c,Attention,GateDownMLP,GateRows,Reduce,DownRows,Affine4,Affine4,Affine4,Affine4),
|
|
366
|
+
d48.stage(w,c,Delta,GateDownMLP,GateRows,Reduce,DownRows,Affine3,Affine3,Affine4,Affine4,Affine3,qk),
|
|
367
|
+
d49.stage(w,c,Delta,GateDownMLP,GateRows,Reduce,DownRows,Affine3,Affine3,Affine4,Affine4,Affine3,qk),
|
|
368
|
+
d50.stage(w,c,Delta,GateDownMLP,GateRows,Reduce,DownK3iRows,Affine3,Affine3,Affine4,Affine4,Affine3,qk),
|
|
369
|
+
a51.stage(w,c,Attention,GateDownMLP,GateRows,Reduce,DownK3iRows,Affine4,Affine4,Affine4,Affine4),
|
|
370
|
+
d52.stage(w,c,Delta,GateDownMLP,GateRows,Reduce,DownK3iRows,Affine3,Affine3,Affine4,Affine4,Affine3,qk),
|
|
371
|
+
d53.stage(w,c,Delta,GateDownMLP,GateRows,Reduce,DownK3iRows,Affine3,Affine4,Affine4,Affine4,Affine3,qk),
|
|
372
|
+
d54.stage(w,c,Delta,GateDownMLP,GateRows,Reduce,DownRows,Affine4,Affine4,Affine4,Affine4,Affine4,qk),
|
|
373
|
+
a55.stage(w,c,Attention,GateDownMLP,GateRows,Reduce,DownRows,Affine4,Affine4,Affine4,Affine4),
|
|
374
|
+
d56.stage(w,c,Delta,GateDownMLP,GateRows,Reduce,DownRows,Affine4,Affine4,Affine4,Affine4,Affine4,qk),
|
|
375
|
+
d57.stage(w,c,Delta,GateDownMLP,GateRows,Reduce,DownRows,Affine4,Affine4,Affine4,Affine4,Affine4,qk),
|
|
376
|
+
d58.stage(w,c,Delta,GateDownMLP,GateRows,Reduce,DownRows,Affine4,Affine3,Affine4,Affine4,Affine3,qk),
|
|
377
|
+
a59.stage(w,c,Attention,GateDownMLP,GateRows,Reduce,DownRows,Affine4,Affine4,Affine4,Affine4),
|
|
378
|
+
d60.stage(w,c,Delta,GateDownMLP,GateRows,Reduce,DownRows,Affine4,Affine4,Affine4,Affine4,Affine3,qk),
|
|
379
|
+
d61.stage(w,c,Delta,GateDownMLP,GateRows,Reduce,DownRows,Affine4,Affine4,Affine4,Affine4,Affine4,qk),
|
|
380
|
+
d62.stage(w,c,Delta,GateDownMLP,GateRows,Reduce,DownK3iRows,Affine4,Affine4,Affine4,Affine4,Affine3,qk),
|
|
381
|
+
a63.stage(w,c,Attention,GateDownMLP,GateRows,Reduce,DownK3iRows,Affine4,Affine4,Affine4,Affine4),
|
|
382
|
+
];
|
|
383
|
+
const rowLayers=this.#rows;
|
|
384
|
+
this.#mma=[
|
|
385
|
+
d0.withProjections(w,c,Delta,rowLayers[0].mlp,AffineMma,AffineMma,AffineMma,AffineMma,AffineMma,qk),
|
|
386
|
+
d1.withProjections(w,c,Delta,rowLayers[1].mlp,AffineMma,AffineMma,AffineMma,AffineMma,AffineMma,qk),
|
|
387
|
+
d2.withProjections(w,c,Delta,rowLayers[2].mlp,AffineMma,AffineMma,AffineMma,AffineMma,AffineMma,qk),
|
|
388
|
+
a3.withProjections(w,c,Attention,rowLayers[3].mlp,AffineMma,AffineMma,AffineMma,AffineMma),
|
|
389
|
+
d4.withProjections(w,c,Delta,rowLayers[4].mlp,AffineMma,AffineMma,AffineMma,AffineMma,AffineMma,qk),
|
|
390
|
+
d5.withProjections(w,c,Delta,rowLayers[5].mlp,AffineMma,AffineMma,AffineMma,AffineMma,AffineMma,qk),
|
|
391
|
+
d6.withProjections(w,c,Delta,rowLayers[6].mlp,AffineMma,AffineMma,AffineMma,AffineMma,AffineMma,qk),
|
|
392
|
+
a7.withProjections(w,c,Attention,rowLayers[7].mlp,AffineMma,AffineMma,AffineMma,AffineMma),
|
|
393
|
+
d8.withProjections(w,c,Delta,rowLayers[8].mlp,AffineMma,AffineMma,AffineMma,AffineMma,AffineMma,qk),
|
|
394
|
+
d9.withProjections(w,c,Delta,rowLayers[9].mlp,AffineMma,AffineMma,AffineMma,AffineMma,AffineMma,qk),
|
|
395
|
+
d10.withProjections(w,c,Delta,rowLayers[10].mlp,AffineMma,AffineMma,AffineMma,AffineMma,AffineMma,qk),
|
|
396
|
+
a11.withProjections(w,c,Attention,rowLayers[11].mlp,AffineMma,AffineMma,AffineMma,AffineMma),
|
|
397
|
+
d12.withProjections(w,c,Delta,rowLayers[12].mlp,AffineMma,AffineMma,AffineMma,AffineMma,AffineMma,qk),
|
|
398
|
+
d13.withProjections(w,c,Delta,rowLayers[13].mlp,AffineMma,AffineMma,AffineMma,AffineMma,AffineMma,qk),
|
|
399
|
+
d14.withProjections(w,c,Delta,rowLayers[14].mlp,AffineMma,AffineMma,AffineMma,AffineMma,AffineMma,qk),
|
|
400
|
+
a15.withProjections(w,c,Attention,rowLayers[15].mlp,AffineMma,AffineMma,AffineMma,AffineMma),
|
|
401
|
+
d16.withProjections(w,c,Delta,rowLayers[16].mlp,AffineMma,AffineMma,AffineMma,AffineMma,AffineMma,qk),
|
|
402
|
+
d17.withProjections(w,c,Delta,rowLayers[17].mlp,AffineMma,AffineMma,AffineMma,AffineMma,AffineMma,qk),
|
|
403
|
+
d18.withProjections(w,c,Delta,rowLayers[18].mlp,AffineMma,AffineMma,AffineMma,AffineMma,AffineMma,qk),
|
|
404
|
+
a19.withProjections(w,c,Attention,rowLayers[19].mlp,AffineMma,AffineMma,AffineMma,AffineMma),
|
|
405
|
+
d20.withProjections(w,c,Delta,rowLayers[20].mlp,AffineMma,AffineMma,AffineMma,AffineMma,AffineMma,qk),
|
|
406
|
+
d21.withProjections(w,c,Delta,rowLayers[21].mlp,AffineMma,AffineMma,AffineMma,AffineMma,AffineMma,qk),
|
|
407
|
+
d22.withProjections(w,c,Delta,rowLayers[22].mlp,AffineMma,AffineMma,AffineMma,AffineMma,AffineMma,qk),
|
|
408
|
+
a23.withProjections(w,c,Attention,rowLayers[23].mlp,AffineMma,AffineMma,AffineMma,AffineMma),
|
|
409
|
+
d24.withProjections(w,c,Delta,rowLayers[24].mlp,AffineMma,AffineMma,AffineMma,AffineMma,AffineMma,qk),
|
|
410
|
+
d25.withProjections(w,c,Delta,rowLayers[25].mlp,AffineMma,AffineMma,AffineMma,AffineMma,AffineMma,qk),
|
|
411
|
+
d26.withProjections(w,c,Delta,rowLayers[26].mlp,AffineMma,AffineMma,AffineMma,AffineMma,AffineMma,qk),
|
|
412
|
+
a27.withProjections(w,c,Attention,rowLayers[27].mlp,AffineMma,AffineMma,AffineMma,AffineMma),
|
|
413
|
+
d28.withProjections(w,c,Delta,rowLayers[28].mlp,AffineMma,AffineMma,AffineMma,AffineMma,AffineMma,qk),
|
|
414
|
+
d29.withProjections(w,c,Delta,rowLayers[29].mlp,AffineMma,AffineMma,AffineMma,AffineMma,AffineMma,qk),
|
|
415
|
+
d30.withProjections(w,c,Delta,rowLayers[30].mlp,AffineMma,AffineMma,AffineMma,AffineMma,AffineMma,qk),
|
|
416
|
+
a31.withProjections(w,c,Attention,rowLayers[31].mlp,AffineMma,AffineMma,AffineMma,AffineMma),
|
|
417
|
+
d32.withProjections(w,c,Delta,rowLayers[32].mlp,AffineMma,AffineMma,AffineMma,AffineMma,AffineMma,qk),
|
|
418
|
+
d33.withProjections(w,c,Delta,rowLayers[33].mlp,AffineMma,AffineMma,AffineMma,AffineMma,AffineMma,qk),
|
|
419
|
+
d34.withProjections(w,c,Delta,rowLayers[34].mlp,AffineMma,AffineMma,AffineMma,AffineMma,AffineMma,qk),
|
|
420
|
+
a35.withProjections(w,c,Attention,rowLayers[35].mlp,AffineMma,AffineMma,AffineMma,AffineMma),
|
|
421
|
+
d36.withProjections(w,c,Delta,rowLayers[36].mlp,AffineMma,AffineMma,AffineMma,AffineMma,AffineMma,qk),
|
|
422
|
+
d37.withProjections(w,c,Delta,rowLayers[37].mlp,AffineMma,AffineMma,AffineMma,AffineMma,AffineMma,qk),
|
|
423
|
+
d38.withProjections(w,c,Delta,rowLayers[38].mlp,AffineMma,AffineMma,AffineMma,AffineMma,AffineMma,qk),
|
|
424
|
+
a39.withProjections(w,c,Attention,rowLayers[39].mlp,AffineMma,AffineMma,AffineMma,AffineMma),
|
|
425
|
+
d40.withProjections(w,c,Delta,rowLayers[40].mlp,AffineMma,AffineMma,AffineMma,AffineMma,AffineMma,qk),
|
|
426
|
+
d41.withProjections(w,c,Delta,rowLayers[41].mlp,AffineMma,AffineMma,AffineMma,AffineMma,AffineMma,qk),
|
|
427
|
+
d42.withProjections(w,c,Delta,rowLayers[42].mlp,AffineMma,AffineMma,AffineMma,AffineMma,AffineMma,qk),
|
|
428
|
+
a43.withProjections(w,c,Attention,rowLayers[43].mlp,AffineMma,AffineMma,AffineMma,AffineMma),
|
|
429
|
+
d44.withProjections(w,c,Delta,rowLayers[44].mlp,AffineMma,AffineMma,AffineMma,AffineMma,AffineMma,qk),
|
|
430
|
+
d45.withProjections(w,c,Delta,rowLayers[45].mlp,AffineMma,AffineMma,AffineMma,AffineMma,AffineMma,qk),
|
|
431
|
+
d46.withProjections(w,c,Delta,rowLayers[46].mlp,AffineMma,AffineMma,AffineMma,AffineMma,AffineMma,qk),
|
|
432
|
+
a47.withProjections(w,c,Attention,rowLayers[47].mlp,AffineMma,AffineMma,AffineMma,AffineMma),
|
|
433
|
+
d48.withProjections(w,c,Delta,rowLayers[48].mlp,AffineMma,AffineMma,AffineMma,AffineMma,AffineMma,qk),
|
|
434
|
+
d49.withProjections(w,c,Delta,rowLayers[49].mlp,AffineMma,AffineMma,AffineMma,AffineMma,AffineMma,qk),
|
|
435
|
+
d50.withProjections(w,c,Delta,rowLayers[50].mlp,AffineMma,AffineMma,AffineMma,AffineMma,AffineMma,qk),
|
|
436
|
+
a51.withProjections(w,c,Attention,rowLayers[51].mlp,AffineMma,AffineMma,AffineMma,AffineMma),
|
|
437
|
+
d52.withProjections(w,c,Delta,rowLayers[52].mlp,AffineMma,AffineMma,AffineMma,AffineMma,AffineMma,qk),
|
|
438
|
+
d53.withProjections(w,c,Delta,rowLayers[53].mlp,AffineMma,AffineMma,AffineMma,AffineMma,AffineMma,qk),
|
|
439
|
+
d54.withProjections(w,c,Delta,rowLayers[54].mlp,AffineMma,AffineMma,AffineMma,AffineMma,AffineMma,qk),
|
|
440
|
+
a55.withProjections(w,c,Attention,rowLayers[55].mlp,AffineMma,AffineMma,AffineMma,AffineMma),
|
|
441
|
+
d56.withProjections(w,c,Delta,rowLayers[56].mlp,AffineMma,AffineMma,AffineMma,AffineMma,AffineMma,qk),
|
|
442
|
+
d57.withProjections(w,c,Delta,rowLayers[57].mlp,AffineMma,AffineMma,AffineMma,AffineMma,AffineMma,qk),
|
|
443
|
+
d58.withProjections(w,c,Delta,rowLayers[58].mlp,AffineMma,AffineMma,AffineMma,AffineMma,AffineMma,qk),
|
|
444
|
+
a59.withProjections(w,c,Attention,rowLayers[59].mlp,AffineMma,AffineMma,AffineMma,AffineMma),
|
|
445
|
+
d60.withProjections(w,c,Delta,rowLayers[60].mlp,AffineMma,AffineMma,AffineMma,AffineMma,AffineMma,qk),
|
|
446
|
+
d61.withProjections(w,c,Delta,rowLayers[61].mlp,AffineMma,AffineMma,AffineMma,AffineMma,AffineMma,qk),
|
|
447
|
+
d62.withProjections(w,c,Delta,rowLayers[62].mlp,AffineMma,AffineMma,AffineMma,AffineMma,AffineMma,qk),
|
|
448
|
+
a63.withProjections(w,c,Attention,rowLayers[63].mlp,AffineMma,AffineMma,AffineMma,AffineMma),
|
|
449
|
+
];
|
|
450
|
+
this.#mmaFive=[
|
|
451
|
+
d0.stage(w,c,Delta,GateDownMLP,GateRows,Tiled,DownRows,AffineMma,AffineMma,AffineMma,AffineMma,AffineMma,qk),
|
|
452
|
+
d1.stage(w,c,Delta,GateDownMLP,GateRows,Tiled,DownK3iRows,AffineMma,AffineMma,AffineMma,AffineMma,AffineMma,qk),
|
|
453
|
+
d2.stage(w,c,Delta,GateDownMLP,GateRows,Tiled,DownK3iRows,AffineMma,AffineMma,AffineMma,AffineMma,AffineMma,qk),
|
|
454
|
+
a3.stage(w,c,Attention,GateDownMLP,GateRows,Tiled,DownK3iRows,AffineMma,AffineMma,AffineMma,AffineMma),
|
|
455
|
+
d4.stage(w,c,Delta,GateDownMLP,GateRows,Tiled,DownK3iRows,AffineMma,AffineMma,AffineMma,AffineMma,AffineMma,qk),
|
|
456
|
+
d5.stage(w,c,Delta,GateDownMLP,GateRows,Tiled,DownK3iRows,AffineMma,AffineMma,AffineMma,AffineMma,AffineMma,qk),
|
|
457
|
+
d6.stage(w,c,Delta,GateDownMLP,GateRows,Tiled,DownK3iRows,AffineMma,AffineMma,AffineMma,AffineMma,AffineMma,qk),
|
|
458
|
+
a7.stage(w,c,Attention,GateDownMLP,GateRows,Tiled,DownK3iRows,AffineMma,AffineMma,AffineMma,AffineMma),
|
|
459
|
+
d8.stage(w,c,Delta,GateDownMLP,GateRows,Tiled,DownK3iRows,AffineMma,AffineMma,AffineMma,AffineMma,AffineMma,qk),
|
|
460
|
+
d9.stage(w,c,Delta,GateDownMLP,GateRows,Tiled,DownK3iRows,AffineMma,AffineMma,AffineMma,AffineMma,AffineMma,qk),
|
|
461
|
+
d10.stage(w,c,Delta,GateDownMLP,GateRows,Tiled,DownK3iRows,AffineMma,AffineMma,AffineMma,AffineMma,AffineMma,qk),
|
|
462
|
+
a11.stage(w,c,Attention,GateDownMLP,GateRows,Tiled,DownRows,AffineMma,AffineMma,AffineMma,AffineMma),
|
|
463
|
+
d12.stage(w,c,Delta,GateDownMLP,GateRows,Tiled,DownK3iRows,AffineMma,AffineMma,AffineMma,AffineMma,AffineMma,qk),
|
|
464
|
+
d13.stage(w,c,Delta,GateDownMLP,GateRows,Tiled,DownRows,AffineMma,AffineMma,AffineMma,AffineMma,AffineMma,qk),
|
|
465
|
+
d14.stage(w,c,Delta,GateDownMLP,GateRows,Tiled,DownK3iRows,AffineMma,AffineMma,AffineMma,AffineMma,AffineMma,qk),
|
|
466
|
+
a15.stage(w,c,Attention,GateDownMLP,GateRows,Tiled,DownRows,AffineMma,AffineMma,AffineMma,AffineMma),
|
|
467
|
+
d16.stage(w,c,Delta,GateDownMLP,GateRows,Tiled,DownRows,AffineMma,AffineMma,AffineMma,AffineMma,AffineMma,qk),
|
|
468
|
+
d17.stage(w,c,Delta,GateDownMLP,GateRows,Tiled,DownRows,AffineMma,AffineMma,AffineMma,AffineMma,AffineMma,qk),
|
|
469
|
+
d18.stage(w,c,Delta,GateDownMLP,GateRows,Tiled,DownRows,AffineMma,AffineMma,AffineMma,AffineMma,AffineMma,qk),
|
|
470
|
+
a19.stage(w,c,Attention,GateDownMLP,GateRows,Tiled,DownK3iRows,AffineMma,AffineMma,AffineMma,AffineMma),
|
|
471
|
+
d20.stage(w,c,Delta,GateDownMLP,GateRows,Tiled,DownK3iRows,AffineMma,AffineMma,AffineMma,AffineMma,AffineMma,qk),
|
|
472
|
+
d21.stage(w,c,Delta,GateDownMLP,GateRows,Tiled,DownK3iRows,AffineMma,AffineMma,AffineMma,AffineMma,AffineMma,qk),
|
|
473
|
+
d22.stage(w,c,Delta,GateDownMLP,GateRows,Tiled,DownK3iRows,AffineMma,AffineMma,AffineMma,AffineMma,AffineMma,qk),
|
|
474
|
+
a23.stage(w,c,Attention,GateDownMLP,GateRows,Tiled,DownK3iRows,AffineMma,AffineMma,AffineMma,AffineMma),
|
|
475
|
+
d24.stage(w,c,Delta,GateDownMLP,GateRows,Tiled,DownK3iRows,AffineMma,AffineMma,AffineMma,AffineMma,AffineMma,qk),
|
|
476
|
+
d25.stage(w,c,Delta,GateDownMLP,GateRows,Tiled,DownK3iRows,AffineMma,AffineMma,AffineMma,AffineMma,AffineMma,qk),
|
|
477
|
+
d26.stage(w,c,Delta,GateDownMLP,GateRows,Tiled,DownK3iRows,AffineMma,AffineMma,AffineMma,AffineMma,AffineMma,qk),
|
|
478
|
+
a27.stage(w,c,Attention,GateDownMLP,GateRows,Tiled,DownK3iRows,AffineMma,AffineMma,AffineMma,AffineMma),
|
|
479
|
+
d28.stage(w,c,Delta,GateDownMLP,GateRows,Tiled,DownK3iRows,AffineMma,AffineMma,AffineMma,AffineMma,AffineMma,qk),
|
|
480
|
+
d29.stage(w,c,Delta,GateDownMLP,GateRows,Tiled,DownRows,AffineMma,AffineMma,AffineMma,AffineMma,AffineMma,qk),
|
|
481
|
+
d30.stage(w,c,Delta,GateDownMLP,GateRows,Tiled,DownK3iRows,AffineMma,AffineMma,AffineMma,AffineMma,AffineMma,qk),
|
|
482
|
+
a31.stage(w,c,Attention,GateDownMLP,GateRows,Tiled,DownRows,AffineMma,AffineMma,AffineMma,AffineMma),
|
|
483
|
+
d32.stage(w,c,Delta,GateDownMLP,GateRows,Tiled,DownRows,AffineMma,AffineMma,AffineMma,AffineMma,AffineMma,qk),
|
|
484
|
+
d33.stage(w,c,Delta,GateDownMLP,GateRows,Tiled,DownRows,AffineMma,AffineMma,AffineMma,AffineMma,AffineMma,qk),
|
|
485
|
+
d34.stage(w,c,Delta,GateDownMLP,GateRows,Tiled,DownK3iRows,AffineMma,AffineMma,AffineMma,AffineMma,AffineMma,qk),
|
|
486
|
+
a35.stage(w,c,Attention,GateDownMLP,GateRows,Tiled,DownK3iRows,AffineMma,AffineMma,AffineMma,AffineMma),
|
|
487
|
+
d36.stage(w,c,Delta,GateDownMLP,GateRows,Tiled,DownK3iRows,AffineMma,AffineMma,AffineMma,AffineMma,AffineMma,qk),
|
|
488
|
+
d37.stage(w,c,Delta,GateDownMLP,GateRows,Tiled,DownK3iRows,AffineMma,AffineMma,AffineMma,AffineMma,AffineMma,qk),
|
|
489
|
+
d38.stage(w,c,Delta,GateDownMLP,GateRows,Tiled,DownK3iRows,AffineMma,AffineMma,AffineMma,AffineMma,AffineMma,qk),
|
|
490
|
+
a39.stage(w,c,Attention,GateDownMLP,GateRows,Tiled,DownK3iRows,AffineMma,AffineMma,AffineMma,AffineMma),
|
|
491
|
+
d40.stage(w,c,Delta,GateDownMLP,GateRows,Tiled,DownK3iRows,AffineMma,AffineMma,AffineMma,AffineMma,AffineMma,qk),
|
|
492
|
+
d41.stage(w,c,Delta,GateDownMLP,GateRows,Tiled,DownK3iRows,AffineMma,AffineMma,AffineMma,AffineMma,AffineMma,qk),
|
|
493
|
+
d42.stage(w,c,Delta,GateDownMLP,GateRows,Tiled,DownK3iRows,AffineMma,AffineMma,AffineMma,AffineMma,AffineMma,qk),
|
|
494
|
+
a43.stage(w,c,Attention,GateDownMLP,GateRows,Tiled,DownK3iRows,AffineMma,AffineMma,AffineMma,AffineMma),
|
|
495
|
+
d44.stage(w,c,Delta,GateDownMLP,GateRows,Tiled,DownK3iRows,AffineMma,AffineMma,AffineMma,AffineMma,AffineMma,qk),
|
|
496
|
+
d45.stage(w,c,Delta,GateDownMLP,GateRows,Tiled,DownRows,AffineMma,AffineMma,AffineMma,AffineMma,AffineMma,qk),
|
|
497
|
+
d46.stage(w,c,Delta,GateDownMLP,GateRows,Tiled,DownRows,AffineMma,AffineMma,AffineMma,AffineMma,AffineMma,qk),
|
|
498
|
+
a47.stage(w,c,Attention,GateDownMLP,GateRows,Tiled,DownRows,AffineMma,AffineMma,AffineMma,AffineMma),
|
|
499
|
+
d48.stage(w,c,Delta,GateDownMLP,GateRows,Tiled,DownRows,AffineMma,AffineMma,AffineMma,AffineMma,AffineMma,qk),
|
|
500
|
+
d49.stage(w,c,Delta,GateDownMLP,GateRows,Tiled,DownRows,AffineMma,AffineMma,AffineMma,AffineMma,AffineMma,qk),
|
|
501
|
+
d50.stage(w,c,Delta,GateDownMLP,GateRows,Tiled,DownK3iRows,AffineMma,AffineMma,AffineMma,AffineMma,AffineMma,qk),
|
|
502
|
+
a51.stage(w,c,Attention,GateDownMLP,GateRows,Tiled,DownK3iRows,AffineMma,AffineMma,AffineMma,AffineMma),
|
|
503
|
+
d52.stage(w,c,Delta,GateDownMLP,GateRows,Tiled,DownK3iRows,AffineMma,AffineMma,AffineMma,AffineMma,AffineMma,qk),
|
|
504
|
+
d53.stage(w,c,Delta,GateDownMLP,GateRows,Tiled,DownK3iRows,AffineMma,AffineMma,AffineMma,AffineMma,AffineMma,qk),
|
|
505
|
+
d54.stage(w,c,Delta,GateDownMLP,GateRows,Tiled,DownRows,AffineMma,AffineMma,AffineMma,AffineMma,AffineMma,qk),
|
|
506
|
+
a55.stage(w,c,Attention,GateDownMLP,GateRows,Tiled,DownRows,AffineMma,AffineMma,AffineMma,AffineMma),
|
|
507
|
+
d56.stage(w,c,Delta,GateDownMLP,GateRows,Tiled,DownRows,AffineMma,AffineMma,AffineMma,AffineMma,AffineMma,qk),
|
|
508
|
+
d57.stage(w,c,Delta,GateDownMLP,GateRows,Tiled,DownRows,AffineMma,AffineMma,AffineMma,AffineMma,AffineMma,qk),
|
|
509
|
+
d58.stage(w,c,Delta,GateDownMLP,GateRows,Tiled,DownRows,AffineMma,AffineMma,AffineMma,AffineMma,AffineMma,qk),
|
|
510
|
+
a59.stage(w,c,Attention,GateDownMLP,GateRows,Tiled,DownRows,AffineMma,AffineMma,AffineMma,AffineMma),
|
|
511
|
+
d60.stage(w,c,Delta,GateDownMLP,GateRows,Tiled,DownRows,AffineMma,AffineMma,AffineMma,AffineMma,AffineMma,qk),
|
|
512
|
+
d61.stage(w,c,Delta,GateDownMLP,GateRows,Tiled,DownRows,AffineMma,AffineMma,AffineMma,AffineMma,AffineMma,qk),
|
|
513
|
+
d62.stage(w,c,Delta,GateDownMLP,GateRows,Tiled,DownK3iRows,AffineMma,AffineMma,AffineMma,AffineMma,AffineMma,qk),
|
|
514
|
+
a63.stage(w,c,Attention,GateDownMLP,GateRows,Tiled,DownK3iRows,AffineMma,AffineMma,AffineMma,AffineMma),
|
|
515
|
+
];
|
|
516
|
+
this.#wide=[
|
|
517
|
+
d0.stage(w,c,Delta,GateDownMLP,WideGate,ExpandGate,WideDown,AffineMma,AffineMma,AffineMma,AffineMma,AffineMma,qk),
|
|
518
|
+
d1.stage(w,c,Delta,GateDownMLP,WideGate,ExpandGate,WideDownK3i,AffineMma,AffineMma,AffineMma,AffineMma,AffineMma,qk),
|
|
519
|
+
d2.stage(w,c,Delta,GateDownMLP,WideGate,ExpandGate,WideDownK3i,AffineMma,AffineMma,AffineMma,AffineMma,AffineMma,qk),
|
|
520
|
+
a3.stage(w,c,Attention,GateDownMLP,WideGate,ExpandGate,WideDownK3i,AffineMma,AffineMma,AffineMma,AffineMma),
|
|
521
|
+
d4.stage(w,c,Delta,GateDownMLP,WideGate,ExpandGate,WideDownK3i,AffineMma,AffineMma,AffineMma,AffineMma,AffineMma,qk),
|
|
522
|
+
d5.stage(w,c,Delta,GateDownMLP,WideGate,ExpandGate,WideDownK3i,AffineMma,AffineMma,AffineMma,AffineMma,AffineMma,qk),
|
|
523
|
+
d6.stage(w,c,Delta,GateDownMLP,WideGate,ExpandGate,WideDownK3i,AffineMma,AffineMma,AffineMma,AffineMma,AffineMma,qk),
|
|
524
|
+
a7.stage(w,c,Attention,GateDownMLP,WideGate,ExpandGate,WideDownK3i,AffineMma,AffineMma,AffineMma,AffineMma),
|
|
525
|
+
d8.stage(w,c,Delta,GateDownMLP,WideGate,ExpandGate,WideDownK3i,AffineMma,AffineMma,AffineMma,AffineMma,AffineMma,qk),
|
|
526
|
+
d9.stage(w,c,Delta,GateDownMLP,WideGate,ExpandGate,WideDownK3i,AffineMma,AffineMma,AffineMma,AffineMma,AffineMma,qk),
|
|
527
|
+
d10.stage(w,c,Delta,GateDownMLP,WideGate,ExpandGate,WideDownK3i,AffineMma,AffineMma,AffineMma,AffineMma,AffineMma,qk),
|
|
528
|
+
a11.stage(w,c,Attention,GateDownMLP,WideGate,ExpandGate,WideDown,AffineMma,AffineMma,AffineMma,AffineMma),
|
|
529
|
+
d12.stage(w,c,Delta,GateDownMLP,WideGate,ExpandGate,WideDownK3i,AffineMma,AffineMma,AffineMma,AffineMma,AffineMma,qk),
|
|
530
|
+
d13.stage(w,c,Delta,GateDownMLP,WideGate,ExpandGate,WideDown,AffineMma,AffineMma,AffineMma,AffineMma,AffineMma,qk),
|
|
531
|
+
d14.stage(w,c,Delta,GateDownMLP,WideGate,ExpandGate,WideDownK3i,AffineMma,AffineMma,AffineMma,AffineMma,AffineMma,qk),
|
|
532
|
+
a15.stage(w,c,Attention,GateDownMLP,WideGate,ExpandGate,WideDown,AffineMma,AffineMma,AffineMma,AffineMma),
|
|
533
|
+
d16.stage(w,c,Delta,GateDownMLP,WideGate,ExpandGate,WideDown,AffineMma,AffineMma,AffineMma,AffineMma,AffineMma,qk),
|
|
534
|
+
d17.stage(w,c,Delta,GateDownMLP,WideGate,ExpandGate,WideDown,AffineMma,AffineMma,AffineMma,AffineMma,AffineMma,qk),
|
|
535
|
+
d18.stage(w,c,Delta,GateDownMLP,WideGate,ExpandGate,WideDown,AffineMma,AffineMma,AffineMma,AffineMma,AffineMma,qk),
|
|
536
|
+
a19.stage(w,c,Attention,GateDownMLP,WideGate,ExpandGate,WideDownK3i,AffineMma,AffineMma,AffineMma,AffineMma),
|
|
537
|
+
d20.stage(w,c,Delta,GateDownMLP,WideGate,ExpandGate,WideDownK3i,AffineMma,AffineMma,AffineMma,AffineMma,AffineMma,qk),
|
|
538
|
+
d21.stage(w,c,Delta,GateDownMLP,WideGate,ExpandGate,WideDownK3i,AffineMma,AffineMma,AffineMma,AffineMma,AffineMma,qk),
|
|
539
|
+
d22.stage(w,c,Delta,GateDownMLP,WideGate,ExpandGate,WideDownK3i,AffineMma,AffineMma,AffineMma,AffineMma,AffineMma,qk),
|
|
540
|
+
a23.stage(w,c,Attention,GateDownMLP,WideGate,ExpandGate,WideDownK3i,AffineMma,AffineMma,AffineMma,AffineMma),
|
|
541
|
+
d24.stage(w,c,Delta,GateDownMLP,WideGate,ExpandGate,WideDownK3i,AffineMma,AffineMma,AffineMma,AffineMma,AffineMma,qk),
|
|
542
|
+
d25.stage(w,c,Delta,GateDownMLP,WideGate,ExpandGate,WideDownK3i,AffineMma,AffineMma,AffineMma,AffineMma,AffineMma,qk),
|
|
543
|
+
d26.stage(w,c,Delta,GateDownMLP,WideGate,ExpandGate,WideDownK3i,AffineMma,AffineMma,AffineMma,AffineMma,AffineMma,qk),
|
|
544
|
+
a27.stage(w,c,Attention,GateDownMLP,WideGate,ExpandGate,WideDownK3i,AffineMma,AffineMma,AffineMma,AffineMma),
|
|
545
|
+
d28.stage(w,c,Delta,GateDownMLP,WideGate,ExpandGate,WideDownK3i,AffineMma,AffineMma,AffineMma,AffineMma,AffineMma,qk),
|
|
546
|
+
d29.stage(w,c,Delta,GateDownMLP,WideGate,ExpandGate,WideDown,AffineMma,AffineMma,AffineMma,AffineMma,AffineMma,qk),
|
|
547
|
+
d30.stage(w,c,Delta,GateDownMLP,WideGate,ExpandGate,WideDownK3i,AffineMma,AffineMma,AffineMma,AffineMma,AffineMma,qk),
|
|
548
|
+
a31.stage(w,c,Attention,GateDownMLP,WideGate,ExpandGate,WideDown,AffineMma,AffineMma,AffineMma,AffineMma),
|
|
549
|
+
d32.stage(w,c,Delta,GateDownMLP,WideGate,ExpandGate,WideDown,AffineMma,AffineMma,AffineMma,AffineMma,AffineMma,qk),
|
|
550
|
+
d33.stage(w,c,Delta,GateDownMLP,WideGate,ExpandGate,WideDown,AffineMma,AffineMma,AffineMma,AffineMma,AffineMma,qk),
|
|
551
|
+
d34.stage(w,c,Delta,GateDownMLP,WideGate,ExpandGate,WideDownK3i,AffineMma,AffineMma,AffineMma,AffineMma,AffineMma,qk),
|
|
552
|
+
a35.stage(w,c,Attention,GateDownMLP,WideGate,ExpandGate,WideDownK3i,AffineMma,AffineMma,AffineMma,AffineMma),
|
|
553
|
+
d36.stage(w,c,Delta,GateDownMLP,WideGate,ExpandGate,WideDownK3i,AffineMma,AffineMma,AffineMma,AffineMma,AffineMma,qk),
|
|
554
|
+
d37.stage(w,c,Delta,GateDownMLP,WideGate,ExpandGate,WideDownK3i,AffineMma,AffineMma,AffineMma,AffineMma,AffineMma,qk),
|
|
555
|
+
d38.stage(w,c,Delta,GateDownMLP,WideGate,ExpandGate,WideDownK3i,AffineMma,AffineMma,AffineMma,AffineMma,AffineMma,qk),
|
|
556
|
+
a39.stage(w,c,Attention,GateDownMLP,WideGate,ExpandGate,WideDownK3i,AffineMma,AffineMma,AffineMma,AffineMma),
|
|
557
|
+
d40.stage(w,c,Delta,GateDownMLP,WideGate,ExpandGate,WideDownK3i,AffineMma,AffineMma,AffineMma,AffineMma,AffineMma,qk),
|
|
558
|
+
d41.stage(w,c,Delta,GateDownMLP,WideGate,ExpandGate,WideDownK3i,AffineMma,AffineMma,AffineMma,AffineMma,AffineMma,qk),
|
|
559
|
+
d42.stage(w,c,Delta,GateDownMLP,WideGate,ExpandGate,WideDownK3i,AffineMma,AffineMma,AffineMma,AffineMma,AffineMma,qk),
|
|
560
|
+
a43.stage(w,c,Attention,GateDownMLP,WideGate,ExpandGate,WideDownK3i,AffineMma,AffineMma,AffineMma,AffineMma),
|
|
561
|
+
d44.stage(w,c,Delta,GateDownMLP,WideGate,ExpandGate,WideDownK3i,AffineMma,AffineMma,AffineMma,AffineMma,AffineMma,qk),
|
|
562
|
+
d45.stage(w,c,Delta,GateDownMLP,WideGate,ExpandGate,WideDown,AffineMma,AffineMma,AffineMma,AffineMma,AffineMma,qk),
|
|
563
|
+
d46.stage(w,c,Delta,GateDownMLP,WideGate,ExpandGate,WideDown,AffineMma,AffineMma,AffineMma,AffineMma,AffineMma,qk),
|
|
564
|
+
a47.stage(w,c,Attention,GateDownMLP,WideGate,ExpandGate,WideDown,AffineMma,AffineMma,AffineMma,AffineMma),
|
|
565
|
+
d48.stage(w,c,Delta,GateDownMLP,WideGate,ExpandGate,WideDown,AffineMma,AffineMma,AffineMma,AffineMma,AffineMma,qk),
|
|
566
|
+
d49.stage(w,c,Delta,GateDownMLP,WideGate,ExpandGate,WideDown,AffineMma,AffineMma,AffineMma,AffineMma,AffineMma,qk),
|
|
567
|
+
d50.stage(w,c,Delta,GateDownMLP,WideGate,ExpandGate,WideDownK3i,AffineMma,AffineMma,AffineMma,AffineMma,AffineMma,qk),
|
|
568
|
+
a51.stage(w,c,Attention,GateDownMLP,WideGate,ExpandGate,WideDownK3i,AffineMma,AffineMma,AffineMma,AffineMma),
|
|
569
|
+
d52.stage(w,c,Delta,GateDownMLP,WideGate,ExpandGate,WideDownK3i,AffineMma,AffineMma,AffineMma,AffineMma,AffineMma,qk),
|
|
570
|
+
d53.stage(w,c,Delta,GateDownMLP,WideGate,ExpandGate,WideDownK3i,AffineMma,AffineMma,AffineMma,AffineMma,AffineMma,qk),
|
|
571
|
+
d54.stage(w,c,Delta,GateDownMLP,WideGate,ExpandGate,WideDown,AffineMma,AffineMma,AffineMma,AffineMma,AffineMma,qk),
|
|
572
|
+
a55.stage(w,c,Attention,GateDownMLP,WideGate,ExpandGate,WideDown,AffineMma,AffineMma,AffineMma,AffineMma),
|
|
573
|
+
d56.stage(w,c,Delta,GateDownMLP,WideGate,ExpandGate,WideDown,AffineMma,AffineMma,AffineMma,AffineMma,AffineMma,qk),
|
|
574
|
+
d57.stage(w,c,Delta,GateDownMLP,WideGate,ExpandGate,WideDown,AffineMma,AffineMma,AffineMma,AffineMma,AffineMma,qk),
|
|
575
|
+
d58.stage(w,c,Delta,GateDownMLP,WideGate,ExpandGate,WideDown,AffineMma,AffineMma,AffineMma,AffineMma,AffineMma,qk),
|
|
576
|
+
a59.stage(w,c,Attention,GateDownMLP,WideGate,ExpandGate,WideDown,AffineMma,AffineMma,AffineMma,AffineMma),
|
|
577
|
+
d60.stage(w,c,Delta,GateDownMLP,WideGate,ExpandGate,WideDown,AffineMma,AffineMma,AffineMma,AffineMma,AffineMma,qk),
|
|
578
|
+
d61.stage(w,c,Delta,GateDownMLP,WideGate,ExpandGate,WideDown,AffineMma,AffineMma,AffineMma,AffineMma,AffineMma,qk),
|
|
579
|
+
d62.stage(w,c,Delta,GateDownMLP,WideGate,ExpandGate,WideDownK3i,AffineMma,AffineMma,AffineMma,AffineMma,AffineMma,qk),
|
|
580
|
+
a63.stage(w,c,Attention,GateDownMLP,WideGate,ExpandGate,WideDownK3i,AffineMma,AffineMma,AffineMma,AffineMma),
|
|
581
|
+
];
|
|
582
|
+
this.#stockNarrow=[
|
|
583
|
+
d0.fused(w,c,Delta,TrellisFusedMLP,Scatter,Plain,qk),
|
|
584
|
+
d1.fused(w,c,Delta,TrellisFusedMLP,Scatter,Plain,qk),
|
|
585
|
+
d2.fused(w,c,Delta,TrellisFusedMLP,Scatter,Plain,qk),
|
|
586
|
+
a3.fused(w,c,Attention,TrellisFusedMLP,Scatter,Plain),
|
|
587
|
+
d4.fused(w,c,Delta,TrellisFusedMLP,Scatter,Plain,qk),
|
|
588
|
+
d5.fused(w,c,Delta,TrellisFusedMLP,Scatter,Plain,qk),
|
|
589
|
+
d6.fused(w,c,Delta,TrellisFusedMLP,Scatter,Plain,qk),
|
|
590
|
+
a7.fused(w,c,Attention,TrellisFusedMLP,Scatter,Plain),
|
|
591
|
+
d8.fused(w,c,Delta,TrellisFusedMLP,Scatter,Plain,qk),
|
|
592
|
+
d9.fused(w,c,Delta,TrellisFusedMLP,Scatter,Plain,qk),
|
|
593
|
+
d10.fused(w,c,Delta,TrellisFusedMLP,Scatter,Plain,qk),
|
|
594
|
+
a11.fused(w,c,Attention,TrellisFusedMLP,Scatter,Plain),
|
|
595
|
+
d12.fused(w,c,Delta,TrellisFusedMLP,Scatter,Plain,qk),
|
|
596
|
+
d13.fused(w,c,Delta,TrellisFusedMLP,Scatter,Plain,qk),
|
|
597
|
+
d14.fused(w,c,Delta,TrellisFusedMLP,Scatter,Plain,qk),
|
|
598
|
+
a15.stage(w,c,Attention,GateDownMLP,MixedGate,Reduce,Scatter,Plain,Plain,Plain,Plain),
|
|
599
|
+
d16.fused(w,c,Delta,TrellisFusedMLP,Scatter,Plain,qk),
|
|
600
|
+
d17.fused(w,c,Delta,TrellisFusedMLP,Scatter,Plain,qk),
|
|
601
|
+
d18.fused(w,c,Delta,TrellisFusedMLP,Scatter,Plain,qk),
|
|
602
|
+
a19.fused(w,c,Attention,TrellisFusedMLP,Scatter,Plain),
|
|
603
|
+
d20.fused(w,c,Delta,TrellisFusedMLP,Scatter,Plain,qk),
|
|
604
|
+
d21.stage(w,c,Delta,GateDownMLP,MixedGate,Reduce,Scatter,Plain,Plain,Plain,Plain,Plain,qk),
|
|
605
|
+
d22.fused(w,c,Delta,TrellisFusedMLP,Scatter,Plain,qk),
|
|
606
|
+
a23.fused(w,c,Attention,TrellisFusedMLP,Scatter,Plain),
|
|
607
|
+
d24.fused(w,c,Delta,TrellisFusedMLP,Scatter,Plain,qk),
|
|
608
|
+
d25.fused(w,c,Delta,TrellisFusedMLP,Scatter,Plain,qk),
|
|
609
|
+
d26.fused(w,c,Delta,TrellisFusedMLP,Scatter,Plain,qk),
|
|
610
|
+
a27.fused(w,c,Attention,TrellisFusedMLP,Scatter,Plain),
|
|
611
|
+
d28.fused(w,c,Delta,TrellisFusedMLP,Scatter,Plain,qk),
|
|
612
|
+
d29.fused(w,c,Delta,TrellisFusedMLP,Scatter,Plain,qk),
|
|
613
|
+
d30.fused(w,c,Delta,TrellisFusedMLP,Scatter,Plain,qk),
|
|
614
|
+
a31.fused(w,c,Attention,TrellisFusedMLP,Scatter,Plain),
|
|
615
|
+
d32.fused(w,c,Delta,TrellisFusedMLP,Scatter,Plain,qk),
|
|
616
|
+
d33.fused(w,c,Delta,TrellisFusedMLP,Scatter,Plain,qk),
|
|
617
|
+
d34.fused(w,c,Delta,TrellisFusedMLP,Scatter,Plain,qk),
|
|
618
|
+
a35.stage(w,c,Attention,GateDownMLP,MixedGate,Reduce,Scatter,Plain,Plain,Plain,Plain),
|
|
619
|
+
d36.stage(w,c,Delta,GateDownMLP,MixedGate,Reduce,Scatter,Plain,Plain,Plain,Plain,Plain,qk),
|
|
620
|
+
d37.fused(w,c,Delta,TrellisFusedMLP,Scatter,Plain,qk),
|
|
621
|
+
d38.fused(w,c,Delta,TrellisFusedMLP,Scatter,Plain,qk),
|
|
622
|
+
a39.fused(w,c,Attention,TrellisFusedMLP,Scatter,Plain),
|
|
623
|
+
d40.fused(w,c,Delta,TrellisFusedMLP,Scatter,Plain,qk),
|
|
624
|
+
d41.fused(w,c,Delta,TrellisFusedMLP,Scatter,Plain,qk),
|
|
625
|
+
d42.fused(w,c,Delta,TrellisFusedMLP,Scatter,Plain,qk),
|
|
626
|
+
a43.stage(w,c,Attention,GateDownMLP,MixedGate,Reduce,Scatter,Plain,Plain,Plain,Plain),
|
|
627
|
+
d44.fused(w,c,Delta,TrellisFusedMLP,Scatter,Plain,qk),
|
|
628
|
+
d45.fused(w,c,Delta,TrellisFusedMLP,Scatter,Plain,qk),
|
|
629
|
+
d46.fused(w,c,Delta,TrellisFusedMLP,Scatter,Plain,qk),
|
|
630
|
+
a47.fused(w,c,Attention,TrellisFusedMLP,Scatter,Plain),
|
|
631
|
+
d48.fused(w,c,Delta,TrellisFusedMLP,Scatter,Plain,qk),
|
|
632
|
+
d49.fused(w,c,Delta,TrellisFusedMLP,Scatter,Plain,qk),
|
|
633
|
+
d50.stage(w,c,Delta,GateDownMLP,MixedGate,Reduce,Scatter,Plain,Plain,Plain,Plain,Plain,qk),
|
|
634
|
+
a51.stage(w,c,Attention,GateDownMLP,MixedGate,Reduce,Scatter,Plain,Plain,Plain,Plain),
|
|
635
|
+
d52.fused(w,c,Delta,TrellisFusedMLP,Scatter,Plain,qk),
|
|
636
|
+
d53.stage(w,c,Delta,GateDownMLP,MixedGate,Reduce,Scatter,Plain,Plain,Plain,Plain,Plain,qk),
|
|
637
|
+
d54.fused(w,c,Delta,TrellisFusedMLP,Scatter,Plain,qk),
|
|
638
|
+
a55.fused(w,c,Attention,TrellisFusedMLP,Scatter,Plain),
|
|
639
|
+
d56.stage(w,c,Delta,GateDownMLP,MixedGate,Reduce,Scatter,Plain,Plain,Plain,Plain,Plain,qk),
|
|
640
|
+
d57.fused(w,c,Delta,TrellisFusedMLP,Scatter,Plain,qk),
|
|
641
|
+
d58.fused(w,c,Delta,TrellisFusedMLP,Scatter,Plain,qk),
|
|
642
|
+
a59.fused(w,c,Attention,TrellisFusedMLP,Scatter,Plain),
|
|
643
|
+
d60.fused(w,c,Delta,TrellisFusedMLP,Scatter,Plain,qk),
|
|
644
|
+
d61.fused(w,c,Delta,TrellisFusedMLP,Scatter,Plain,qk),
|
|
645
|
+
d62.stage(w,c,Delta,GateDownMLP,MixedGate,Reduce,Scatter,Plain,Plain,Plain,Plain,Plain,qk),
|
|
646
|
+
a63.fused(w,c,Attention,TrellisFusedMLP,Scatter,Plain),
|
|
647
|
+
];
|
|
648
|
+
this.#fiveToEight=[
|
|
649
|
+
d0.prompt(w,c,Delta,TrellisUnfusedMLP,Tiled,SplitK,Plain,qk),
|
|
650
|
+
d1.prompt(w,c,Delta,TrellisUnfusedMLP,Tiled,SplitK,Plain,qk),
|
|
651
|
+
d2.prompt(w,c,Delta,TrellisUnfusedMLP,Tiled,SplitK,Plain,qk),
|
|
652
|
+
a3.prompt(w,c,Attention,TrellisUnfusedMLP,Tiled,SplitK,Plain),
|
|
653
|
+
d4.prompt(w,c,Delta,TrellisUnfusedMLP,Tiled,SplitK,Plain,qk),
|
|
654
|
+
d5.prompt(w,c,Delta,TrellisUnfusedMLP,Tiled,SplitK,Plain,qk),
|
|
655
|
+
d6.prompt(w,c,Delta,TrellisUnfusedMLP,Tiled,SplitK,Plain,qk),
|
|
656
|
+
a7.prompt(w,c,Attention,TrellisUnfusedMLP,Tiled,SplitK,Plain),
|
|
657
|
+
d8.prompt(w,c,Delta,TrellisUnfusedMLP,Tiled,SplitK,Plain,qk),
|
|
658
|
+
d9.prompt(w,c,Delta,TrellisUnfusedMLP,Tiled,SplitK,Plain,qk),
|
|
659
|
+
d10.prompt(w,c,Delta,TrellisUnfusedMLP,Tiled,SplitK,Plain,qk),
|
|
660
|
+
a11.prompt(w,c,Attention,TrellisUnfusedMLP,Tiled,SplitK,Plain),
|
|
661
|
+
d12.prompt(w,c,Delta,TrellisUnfusedMLP,Tiled,SplitK,Plain,qk),
|
|
662
|
+
d13.prompt(w,c,Delta,TrellisUnfusedMLP,Tiled,SplitK,Plain,qk),
|
|
663
|
+
d14.prompt(w,c,Delta,TrellisUnfusedMLP,Tiled,SplitK,Plain,qk),
|
|
664
|
+
a15.prompt(w,c,Attention,TrellisUnfusedMLP,Tiled,SplitK,Plain),
|
|
665
|
+
d16.prompt(w,c,Delta,TrellisUnfusedMLP,Tiled,SplitK,Plain,qk),
|
|
666
|
+
d17.prompt(w,c,Delta,TrellisUnfusedMLP,Tiled,SplitK,Plain,qk),
|
|
667
|
+
d18.prompt(w,c,Delta,TrellisUnfusedMLP,Tiled,SplitK,Plain,qk),
|
|
668
|
+
a19.prompt(w,c,Attention,TrellisUnfusedMLP,Tiled,SplitK,Plain),
|
|
669
|
+
d20.prompt(w,c,Delta,TrellisUnfusedMLP,Tiled,SplitK,Plain,qk),
|
|
670
|
+
d21.prompt(w,c,Delta,TrellisUnfusedMLP,Tiled,SplitK,Plain,qk),
|
|
671
|
+
d22.prompt(w,c,Delta,TrellisUnfusedMLP,Tiled,SplitK,Plain,qk),
|
|
672
|
+
a23.prompt(w,c,Attention,TrellisUnfusedMLP,Tiled,SplitK,Plain),
|
|
673
|
+
d24.prompt(w,c,Delta,TrellisUnfusedMLP,Tiled,SplitK,Plain,qk),
|
|
674
|
+
d25.prompt(w,c,Delta,TrellisUnfusedMLP,Tiled,SplitK,Plain,qk),
|
|
675
|
+
d26.prompt(w,c,Delta,TrellisUnfusedMLP,Tiled,SplitK,Plain,qk),
|
|
676
|
+
a27.prompt(w,c,Attention,TrellisUnfusedMLP,Tiled,SplitK,Plain),
|
|
677
|
+
d28.prompt(w,c,Delta,TrellisUnfusedMLP,Tiled,SplitK,Plain,qk),
|
|
678
|
+
d29.prompt(w,c,Delta,TrellisUnfusedMLP,Tiled,SplitK,Plain,qk),
|
|
679
|
+
d30.prompt(w,c,Delta,TrellisUnfusedMLP,Tiled,SplitK,Plain,qk),
|
|
680
|
+
a31.prompt(w,c,Attention,TrellisUnfusedMLP,Tiled,SplitK,Plain),
|
|
681
|
+
d32.prompt(w,c,Delta,TrellisUnfusedMLP,Tiled,SplitK,Plain,qk),
|
|
682
|
+
d33.prompt(w,c,Delta,TrellisUnfusedMLP,Tiled,SplitK,Plain,qk),
|
|
683
|
+
d34.prompt(w,c,Delta,TrellisUnfusedMLP,Tiled,SplitK,Plain,qk),
|
|
684
|
+
a35.prompt(w,c,Attention,TrellisUnfusedMLP,Tiled,SplitK,Plain),
|
|
685
|
+
d36.prompt(w,c,Delta,TrellisUnfusedMLP,Tiled,SplitK,Plain,qk),
|
|
686
|
+
d37.prompt(w,c,Delta,TrellisUnfusedMLP,Tiled,SplitK,Plain,qk),
|
|
687
|
+
d38.prompt(w,c,Delta,TrellisUnfusedMLP,Tiled,SplitK,Plain,qk),
|
|
688
|
+
a39.prompt(w,c,Attention,TrellisUnfusedMLP,Tiled,SplitK,Plain),
|
|
689
|
+
d40.prompt(w,c,Delta,TrellisUnfusedMLP,Tiled,SplitK,Plain,qk),
|
|
690
|
+
d41.prompt(w,c,Delta,TrellisUnfusedMLP,Tiled,SplitK,Plain,qk),
|
|
691
|
+
d42.prompt(w,c,Delta,TrellisUnfusedMLP,Tiled,SplitK,Plain,qk),
|
|
692
|
+
a43.prompt(w,c,Attention,TrellisUnfusedMLP,Tiled,SplitK,Plain),
|
|
693
|
+
d44.prompt(w,c,Delta,TrellisUnfusedMLP,Tiled,SplitK,Plain,qk),
|
|
694
|
+
d45.prompt(w,c,Delta,TrellisUnfusedMLP,Tiled,SplitK,Plain,qk),
|
|
695
|
+
d46.prompt(w,c,Delta,TrellisUnfusedMLP,Tiled,SplitK,Plain,qk),
|
|
696
|
+
a47.prompt(w,c,Attention,TrellisUnfusedMLP,Tiled,SplitK,Plain),
|
|
697
|
+
d48.prompt(w,c,Delta,TrellisUnfusedMLP,Tiled,SplitK,Plain,qk),
|
|
698
|
+
d49.prompt(w,c,Delta,TrellisUnfusedMLP,Tiled,SplitK,Plain,qk),
|
|
699
|
+
d50.prompt(w,c,Delta,TrellisUnfusedMLP,Tiled,SplitK,Plain,qk),
|
|
700
|
+
a51.prompt(w,c,Attention,TrellisUnfusedMLP,Tiled,SplitK,Plain),
|
|
701
|
+
d52.prompt(w,c,Delta,TrellisUnfusedMLP,Tiled,SplitK,Plain,qk),
|
|
702
|
+
d53.prompt(w,c,Delta,TrellisUnfusedMLP,Tiled,SplitK,Plain,qk),
|
|
703
|
+
d54.prompt(w,c,Delta,TrellisUnfusedMLP,Tiled,SplitK,Plain,qk),
|
|
704
|
+
a55.prompt(w,c,Attention,TrellisUnfusedMLP,Tiled,SplitK,Plain),
|
|
705
|
+
d56.prompt(w,c,Delta,TrellisUnfusedMLP,Tiled,SplitK,Plain,qk),
|
|
706
|
+
d57.prompt(w,c,Delta,TrellisUnfusedMLP,Tiled,SplitK,Plain,qk),
|
|
707
|
+
d58.prompt(w,c,Delta,TrellisUnfusedMLP,Tiled,SplitK,Plain,qk),
|
|
708
|
+
a59.prompt(w,c,Attention,TrellisUnfusedMLP,Tiled,SplitK,Plain),
|
|
709
|
+
d60.prompt(w,c,Delta,TrellisUnfusedMLP,Tiled,SplitK,Plain,qk),
|
|
710
|
+
d61.prompt(w,c,Delta,TrellisUnfusedMLP,Tiled,SplitK,Plain,qk),
|
|
711
|
+
d62.prompt(w,c,Delta,TrellisUnfusedMLP,Tiled,SplitK,Plain,qk),
|
|
712
|
+
a63.prompt(w,c,Attention,TrellisUnfusedMLP,Tiled,SplitK,Plain),
|
|
713
|
+
];
|
|
714
|
+
this.#tiled=[
|
|
715
|
+
d0.prompt(w,c,Delta,TrellisUnfusedMLP,Tiled,ExpandDown,Plain,qk),
|
|
716
|
+
d1.prompt(w,c,Delta,TrellisUnfusedMLP,Tiled,ExpandDown,Plain,qk),
|
|
717
|
+
d2.prompt(w,c,Delta,TrellisUnfusedMLP,Tiled,ExpandDown,Plain,qk),
|
|
718
|
+
a3.prompt(w,c,Attention,TrellisUnfusedMLP,Tiled,ExpandDown,Plain),
|
|
719
|
+
d4.prompt(w,c,Delta,TrellisUnfusedMLP,Tiled,ExpandDown,Plain,qk),
|
|
720
|
+
d5.prompt(w,c,Delta,TrellisUnfusedMLP,Tiled,ExpandDown,Plain,qk),
|
|
721
|
+
d6.prompt(w,c,Delta,TrellisUnfusedMLP,Tiled,ExpandDown,Plain,qk),
|
|
722
|
+
a7.prompt(w,c,Attention,TrellisUnfusedMLP,Tiled,ExpandDown,Plain),
|
|
723
|
+
d8.prompt(w,c,Delta,TrellisUnfusedMLP,Tiled,ExpandDown,Plain,qk),
|
|
724
|
+
d9.prompt(w,c,Delta,TrellisUnfusedMLP,Tiled,ExpandDown,Plain,qk),
|
|
725
|
+
d10.prompt(w,c,Delta,TrellisUnfusedMLP,Tiled,ExpandDown,Plain,qk),
|
|
726
|
+
a11.prompt(w,c,Attention,TrellisUnfusedMLP,Tiled,ExpandDown,Plain),
|
|
727
|
+
d12.prompt(w,c,Delta,TrellisUnfusedMLP,Tiled,ExpandDown,Plain,qk),
|
|
728
|
+
d13.prompt(w,c,Delta,TrellisUnfusedMLP,Tiled,ExpandDown,Plain,qk),
|
|
729
|
+
d14.prompt(w,c,Delta,TrellisUnfusedMLP,Tiled,ExpandDown,Plain,qk),
|
|
730
|
+
a15.prompt(w,c,Attention,TrellisUnfusedMLP,Tiled,ExpandDown,Plain),
|
|
731
|
+
d16.prompt(w,c,Delta,TrellisUnfusedMLP,Tiled,ExpandDown,Plain,qk),
|
|
732
|
+
d17.prompt(w,c,Delta,TrellisUnfusedMLP,Tiled,ExpandDown,Plain,qk),
|
|
733
|
+
d18.prompt(w,c,Delta,TrellisUnfusedMLP,Tiled,ExpandDown,Plain,qk),
|
|
734
|
+
a19.prompt(w,c,Attention,TrellisUnfusedMLP,Tiled,ExpandDown,Plain),
|
|
735
|
+
d20.prompt(w,c,Delta,TrellisUnfusedMLP,Tiled,ExpandDown,Plain,qk),
|
|
736
|
+
d21.prompt(w,c,Delta,TrellisUnfusedMLP,Tiled,ExpandDown,Plain,qk),
|
|
737
|
+
d22.prompt(w,c,Delta,TrellisUnfusedMLP,Tiled,ExpandDown,Plain,qk),
|
|
738
|
+
a23.prompt(w,c,Attention,TrellisUnfusedMLP,Tiled,ExpandDown,Plain),
|
|
739
|
+
d24.prompt(w,c,Delta,TrellisUnfusedMLP,Tiled,ExpandDown,Plain,qk),
|
|
740
|
+
d25.prompt(w,c,Delta,TrellisUnfusedMLP,Tiled,ExpandDown,Plain,qk),
|
|
741
|
+
d26.prompt(w,c,Delta,TrellisUnfusedMLP,Tiled,ExpandDown,Plain,qk),
|
|
742
|
+
a27.prompt(w,c,Attention,TrellisUnfusedMLP,Tiled,ExpandDown,Plain),
|
|
743
|
+
d28.prompt(w,c,Delta,TrellisUnfusedMLP,Tiled,ExpandDown,Plain,qk),
|
|
744
|
+
d29.prompt(w,c,Delta,TrellisUnfusedMLP,Tiled,ExpandDown,Plain,qk),
|
|
745
|
+
d30.prompt(w,c,Delta,TrellisUnfusedMLP,Tiled,ExpandDown,Plain,qk),
|
|
746
|
+
a31.prompt(w,c,Attention,TrellisUnfusedMLP,Tiled,ExpandDown,Plain),
|
|
747
|
+
d32.prompt(w,c,Delta,TrellisUnfusedMLP,Tiled,ExpandDown,Plain,qk),
|
|
748
|
+
d33.prompt(w,c,Delta,TrellisUnfusedMLP,Tiled,ExpandDown,Plain,qk),
|
|
749
|
+
d34.prompt(w,c,Delta,TrellisUnfusedMLP,Tiled,ExpandDown,Plain,qk),
|
|
750
|
+
a35.prompt(w,c,Attention,TrellisUnfusedMLP,Tiled,ExpandDown,Plain),
|
|
751
|
+
d36.prompt(w,c,Delta,TrellisUnfusedMLP,Tiled,ExpandDown,Plain,qk),
|
|
752
|
+
d37.prompt(w,c,Delta,TrellisUnfusedMLP,Tiled,ExpandDown,Plain,qk),
|
|
753
|
+
d38.prompt(w,c,Delta,TrellisUnfusedMLP,Tiled,ExpandDown,Plain,qk),
|
|
754
|
+
a39.prompt(w,c,Attention,TrellisUnfusedMLP,Tiled,ExpandDown,Plain),
|
|
755
|
+
d40.prompt(w,c,Delta,TrellisUnfusedMLP,Tiled,ExpandDown,Plain,qk),
|
|
756
|
+
d41.prompt(w,c,Delta,TrellisUnfusedMLP,Tiled,ExpandDown,Plain,qk),
|
|
757
|
+
d42.prompt(w,c,Delta,TrellisUnfusedMLP,Tiled,ExpandDown,Plain,qk),
|
|
758
|
+
a43.prompt(w,c,Attention,TrellisUnfusedMLP,Tiled,ExpandDown,Plain),
|
|
759
|
+
d44.prompt(w,c,Delta,TrellisUnfusedMLP,Tiled,ExpandDown,Plain,qk),
|
|
760
|
+
d45.prompt(w,c,Delta,TrellisUnfusedMLP,Tiled,ExpandDown,Plain,qk),
|
|
761
|
+
d46.prompt(w,c,Delta,TrellisUnfusedMLP,Tiled,ExpandDown,Plain,qk),
|
|
762
|
+
a47.prompt(w,c,Attention,TrellisUnfusedMLP,Tiled,ExpandDown,Plain),
|
|
763
|
+
d48.prompt(w,c,Delta,TrellisUnfusedMLP,Tiled,ExpandDown,Plain,qk),
|
|
764
|
+
d49.prompt(w,c,Delta,TrellisUnfusedMLP,Tiled,ExpandDown,Plain,qk),
|
|
765
|
+
d50.prompt(w,c,Delta,TrellisUnfusedMLP,Tiled,ExpandDown,Plain,qk),
|
|
766
|
+
a51.prompt(w,c,Attention,TrellisUnfusedMLP,Tiled,ExpandDown,Plain),
|
|
767
|
+
d52.prompt(w,c,Delta,TrellisUnfusedMLP,Tiled,ExpandDown,Plain,qk),
|
|
768
|
+
d53.prompt(w,c,Delta,TrellisUnfusedMLP,Tiled,ExpandDown,Plain,qk),
|
|
769
|
+
d54.prompt(w,c,Delta,TrellisUnfusedMLP,Tiled,ExpandDown,Plain,qk),
|
|
770
|
+
a55.prompt(w,c,Attention,TrellisUnfusedMLP,Tiled,ExpandDown,Plain),
|
|
771
|
+
d56.prompt(w,c,Delta,TrellisUnfusedMLP,Tiled,ExpandDown,Plain,qk),
|
|
772
|
+
d57.prompt(w,c,Delta,TrellisUnfusedMLP,Tiled,ExpandDown,Plain,qk),
|
|
773
|
+
d58.prompt(w,c,Delta,TrellisUnfusedMLP,Tiled,ExpandDown,Plain,qk),
|
|
774
|
+
a59.prompt(w,c,Attention,TrellisUnfusedMLP,Tiled,ExpandDown,Plain),
|
|
775
|
+
d60.prompt(w,c,Delta,TrellisUnfusedMLP,Tiled,ExpandDown,Plain,qk),
|
|
776
|
+
d61.prompt(w,c,Delta,TrellisUnfusedMLP,Tiled,ExpandDown,Plain,qk),
|
|
777
|
+
d62.prompt(w,c,Delta,TrellisUnfusedMLP,Tiled,ExpandDown,Plain,qk),
|
|
778
|
+
a63.prompt(w,c,Attention,TrellisUnfusedMLP,Tiled,ExpandDown,Plain),
|
|
779
|
+
];
|
|
780
|
+
this.#expanded=[
|
|
781
|
+
d0.prompt(w,c,Delta,TrellisUnfusedMLP,ExpandGate,ExpandDown,Plain,qk),
|
|
782
|
+
d1.prompt(w,c,Delta,TrellisUnfusedMLP,ExpandGate,ExpandDown,Plain,qk),
|
|
783
|
+
d2.prompt(w,c,Delta,TrellisUnfusedMLP,ExpandGate,ExpandDown,Plain,qk),
|
|
784
|
+
a3.prompt(w,c,Attention,TrellisUnfusedMLP,ExpandGate,ExpandDown,Plain),
|
|
785
|
+
d4.prompt(w,c,Delta,TrellisUnfusedMLP,ExpandGate,ExpandDown,Plain,qk),
|
|
786
|
+
d5.prompt(w,c,Delta,TrellisUnfusedMLP,ExpandGate,ExpandDown,Plain,qk),
|
|
787
|
+
d6.prompt(w,c,Delta,TrellisUnfusedMLP,ExpandGate,ExpandDown,Plain,qk),
|
|
788
|
+
a7.prompt(w,c,Attention,TrellisUnfusedMLP,ExpandGate,ExpandDown,Plain),
|
|
789
|
+
d8.prompt(w,c,Delta,TrellisUnfusedMLP,ExpandGate,ExpandDown,Plain,qk),
|
|
790
|
+
d9.prompt(w,c,Delta,TrellisUnfusedMLP,ExpandGate,ExpandDown,Plain,qk),
|
|
791
|
+
d10.prompt(w,c,Delta,TrellisUnfusedMLP,ExpandGate,ExpandDown,Plain,qk),
|
|
792
|
+
a11.prompt(w,c,Attention,TrellisUnfusedMLP,ExpandGate,ExpandDown,Plain),
|
|
793
|
+
d12.prompt(w,c,Delta,TrellisUnfusedMLP,ExpandGate,ExpandDown,Plain,qk),
|
|
794
|
+
d13.prompt(w,c,Delta,TrellisUnfusedMLP,ExpandGate,ExpandDown,Plain,qk),
|
|
795
|
+
d14.prompt(w,c,Delta,TrellisUnfusedMLP,ExpandGate,ExpandDown,Plain,qk),
|
|
796
|
+
a15.prompt(w,c,Attention,TrellisUnfusedMLP,ExpandGate,ExpandDown,Plain),
|
|
797
|
+
d16.prompt(w,c,Delta,TrellisUnfusedMLP,ExpandGate,ExpandDown,Plain,qk),
|
|
798
|
+
d17.prompt(w,c,Delta,TrellisUnfusedMLP,ExpandGate,ExpandDown,Plain,qk),
|
|
799
|
+
d18.prompt(w,c,Delta,TrellisUnfusedMLP,ExpandGate,ExpandDown,Plain,qk),
|
|
800
|
+
a19.prompt(w,c,Attention,TrellisUnfusedMLP,ExpandGate,ExpandDown,Plain),
|
|
801
|
+
d20.prompt(w,c,Delta,TrellisUnfusedMLP,ExpandGate,ExpandDown,Plain,qk),
|
|
802
|
+
d21.prompt(w,c,Delta,TrellisUnfusedMLP,ExpandGate,ExpandDown,Plain,qk),
|
|
803
|
+
d22.prompt(w,c,Delta,TrellisUnfusedMLP,ExpandGate,ExpandDown,Plain,qk),
|
|
804
|
+
a23.prompt(w,c,Attention,TrellisUnfusedMLP,ExpandGate,ExpandDown,Plain),
|
|
805
|
+
d24.prompt(w,c,Delta,TrellisUnfusedMLP,ExpandGate,ExpandDown,Plain,qk),
|
|
806
|
+
d25.prompt(w,c,Delta,TrellisUnfusedMLP,ExpandGate,ExpandDown,Plain,qk),
|
|
807
|
+
d26.prompt(w,c,Delta,TrellisUnfusedMLP,ExpandGate,ExpandDown,Plain,qk),
|
|
808
|
+
a27.prompt(w,c,Attention,TrellisUnfusedMLP,ExpandGate,ExpandDown,Plain),
|
|
809
|
+
d28.prompt(w,c,Delta,TrellisUnfusedMLP,ExpandGate,ExpandDown,Plain,qk),
|
|
810
|
+
d29.prompt(w,c,Delta,TrellisUnfusedMLP,ExpandGate,ExpandDown,Plain,qk),
|
|
811
|
+
d30.prompt(w,c,Delta,TrellisUnfusedMLP,ExpandGate,ExpandDown,Plain,qk),
|
|
812
|
+
a31.prompt(w,c,Attention,TrellisUnfusedMLP,ExpandGate,ExpandDown,Plain),
|
|
813
|
+
d32.prompt(w,c,Delta,TrellisUnfusedMLP,ExpandGate,ExpandDown,Plain,qk),
|
|
814
|
+
d33.prompt(w,c,Delta,TrellisUnfusedMLP,ExpandGate,ExpandDown,Plain,qk),
|
|
815
|
+
d34.prompt(w,c,Delta,TrellisUnfusedMLP,ExpandGate,ExpandDown,Plain,qk),
|
|
816
|
+
a35.prompt(w,c,Attention,TrellisUnfusedMLP,ExpandGate,ExpandDown,Plain),
|
|
817
|
+
d36.prompt(w,c,Delta,TrellisUnfusedMLP,ExpandGate,ExpandDown,Plain,qk),
|
|
818
|
+
d37.prompt(w,c,Delta,TrellisUnfusedMLP,ExpandGate,ExpandDown,Plain,qk),
|
|
819
|
+
d38.prompt(w,c,Delta,TrellisUnfusedMLP,ExpandGate,ExpandDown,Plain,qk),
|
|
820
|
+
a39.prompt(w,c,Attention,TrellisUnfusedMLP,ExpandGate,ExpandDown,Plain),
|
|
821
|
+
d40.prompt(w,c,Delta,TrellisUnfusedMLP,ExpandGate,ExpandDown,Plain,qk),
|
|
822
|
+
d41.prompt(w,c,Delta,TrellisUnfusedMLP,ExpandGate,ExpandDown,Plain,qk),
|
|
823
|
+
d42.prompt(w,c,Delta,TrellisUnfusedMLP,ExpandGate,ExpandDown,Plain,qk),
|
|
824
|
+
a43.prompt(w,c,Attention,TrellisUnfusedMLP,ExpandGate,ExpandDown,Plain),
|
|
825
|
+
d44.prompt(w,c,Delta,TrellisUnfusedMLP,ExpandGate,ExpandDown,Plain,qk),
|
|
826
|
+
d45.prompt(w,c,Delta,TrellisUnfusedMLP,ExpandGate,ExpandDown,Plain,qk),
|
|
827
|
+
d46.prompt(w,c,Delta,TrellisUnfusedMLP,ExpandGate,ExpandDown,Plain,qk),
|
|
828
|
+
a47.prompt(w,c,Attention,TrellisUnfusedMLP,ExpandGate,ExpandDown,Plain),
|
|
829
|
+
d48.prompt(w,c,Delta,TrellisUnfusedMLP,ExpandGate,ExpandDown,Plain,qk),
|
|
830
|
+
d49.prompt(w,c,Delta,TrellisUnfusedMLP,ExpandGate,ExpandDown,Plain,qk),
|
|
831
|
+
d50.prompt(w,c,Delta,TrellisUnfusedMLP,ExpandGate,ExpandDown,Plain,qk),
|
|
832
|
+
a51.prompt(w,c,Attention,TrellisUnfusedMLP,ExpandGate,ExpandDown,Plain),
|
|
833
|
+
d52.prompt(w,c,Delta,TrellisUnfusedMLP,ExpandGate,ExpandDown,Plain,qk),
|
|
834
|
+
d53.prompt(w,c,Delta,TrellisUnfusedMLP,ExpandGate,ExpandDown,Plain,qk),
|
|
835
|
+
d54.prompt(w,c,Delta,TrellisUnfusedMLP,ExpandGate,ExpandDown,Plain,qk),
|
|
836
|
+
a55.prompt(w,c,Attention,TrellisUnfusedMLP,ExpandGate,ExpandDown,Plain),
|
|
837
|
+
d56.prompt(w,c,Delta,TrellisUnfusedMLP,ExpandGate,ExpandDown,Plain,qk),
|
|
838
|
+
d57.prompt(w,c,Delta,TrellisUnfusedMLP,ExpandGate,ExpandDown,Plain,qk),
|
|
839
|
+
d58.prompt(w,c,Delta,TrellisUnfusedMLP,ExpandGate,ExpandDown,Plain,qk),
|
|
840
|
+
a59.prompt(w,c,Attention,TrellisUnfusedMLP,ExpandGate,ExpandDown,Plain),
|
|
841
|
+
d60.prompt(w,c,Delta,TrellisUnfusedMLP,ExpandGate,ExpandDown,Plain,qk),
|
|
842
|
+
d61.prompt(w,c,Delta,TrellisUnfusedMLP,ExpandGate,ExpandDown,Plain,qk),
|
|
843
|
+
d62.prompt(w,c,Delta,TrellisUnfusedMLP,ExpandGate,ExpandDown,Plain,qk),
|
|
844
|
+
a63.prompt(w,c,Attention,TrellisUnfusedMLP,ExpandGate,ExpandDown,Plain),
|
|
845
|
+
];
|
|
846
|
+
const stock=this.#stockNarrow,five=this.#fiveToEight,tiled=this.#tiled,expanded=this.#expanded;
|
|
847
|
+
const mixedUnfused:Layers=[
|
|
848
|
+
d0.prompt(w,c,Delta,TrellisUnfusedMLP,Reduce,Scatter,Plain,qk),d1.prompt(w,c,Delta,TrellisUnfusedMLP,Reduce,Scatter,Plain,qk),d2.prompt(w,c,Delta,TrellisUnfusedMLP,Reduce,Scatter,Plain,qk),a3.prompt(w,c,Attention,TrellisUnfusedMLP,Reduce,Scatter,Plain),
|
|
849
|
+
d4.prompt(w,c,Delta,TrellisUnfusedMLP,Reduce,Scatter,Plain,qk),d5.prompt(w,c,Delta,TrellisUnfusedMLP,Reduce,Scatter,Plain,qk),d6.prompt(w,c,Delta,TrellisUnfusedMLP,Reduce,Scatter,Plain,qk),a7.prompt(w,c,Attention,TrellisUnfusedMLP,Reduce,Scatter,Plain),
|
|
850
|
+
d8.prompt(w,c,Delta,TrellisUnfusedMLP,Reduce,Scatter,Plain,qk),d9.prompt(w,c,Delta,TrellisUnfusedMLP,Reduce,Scatter,Plain,qk),d10.prompt(w,c,Delta,TrellisUnfusedMLP,Reduce,Scatter,Plain,qk),a11.prompt(w,c,Attention,TrellisUnfusedMLP,Reduce,Scatter,Plain),
|
|
851
|
+
d12.prompt(w,c,Delta,TrellisUnfusedMLP,Reduce,Scatter,Plain,qk),d13.prompt(w,c,Delta,TrellisUnfusedMLP,Reduce,Scatter,Plain,qk),d14.prompt(w,c,Delta,TrellisUnfusedMLP,Reduce,Scatter,Plain,qk),a15.prompt(w,c,Attention,TrellisUnfusedMLP,Reduce,Scatter,Plain),
|
|
852
|
+
d16.prompt(w,c,Delta,TrellisUnfusedMLP,Reduce,Scatter,Plain,qk),d17.prompt(w,c,Delta,TrellisUnfusedMLP,Reduce,Scatter,Plain,qk),d18.prompt(w,c,Delta,TrellisUnfusedMLP,Reduce,Scatter,Plain,qk),a19.prompt(w,c,Attention,TrellisUnfusedMLP,Reduce,Scatter,Plain),
|
|
853
|
+
d20.prompt(w,c,Delta,TrellisUnfusedMLP,Reduce,Scatter,Plain,qk),d21.prompt(w,c,Delta,TrellisUnfusedMLP,Reduce,Scatter,Plain,qk),d22.prompt(w,c,Delta,TrellisUnfusedMLP,Reduce,Scatter,Plain,qk),a23.prompt(w,c,Attention,TrellisUnfusedMLP,Reduce,Scatter,Plain),
|
|
854
|
+
d24.prompt(w,c,Delta,TrellisUnfusedMLP,Reduce,Scatter,Plain,qk),d25.prompt(w,c,Delta,TrellisUnfusedMLP,Reduce,Scatter,Plain,qk),d26.prompt(w,c,Delta,TrellisUnfusedMLP,Reduce,Scatter,Plain,qk),a27.prompt(w,c,Attention,TrellisUnfusedMLP,Reduce,Scatter,Plain),
|
|
855
|
+
d28.prompt(w,c,Delta,TrellisUnfusedMLP,Reduce,Scatter,Plain,qk),d29.prompt(w,c,Delta,TrellisUnfusedMLP,Reduce,Scatter,Plain,qk),d30.prompt(w,c,Delta,TrellisUnfusedMLP,Reduce,Scatter,Plain,qk),a31.prompt(w,c,Attention,TrellisUnfusedMLP,Reduce,Scatter,Plain),
|
|
856
|
+
d32.prompt(w,c,Delta,TrellisUnfusedMLP,Reduce,Scatter,Plain,qk),d33.prompt(w,c,Delta,TrellisUnfusedMLP,Reduce,Scatter,Plain,qk),d34.prompt(w,c,Delta,TrellisUnfusedMLP,Reduce,Scatter,Plain,qk),a35.prompt(w,c,Attention,TrellisUnfusedMLP,Reduce,Scatter,Plain),
|
|
857
|
+
d36.prompt(w,c,Delta,TrellisUnfusedMLP,Reduce,Scatter,Plain,qk),d37.prompt(w,c,Delta,TrellisUnfusedMLP,Reduce,Scatter,Plain,qk),d38.prompt(w,c,Delta,TrellisUnfusedMLP,Reduce,Scatter,Plain,qk),a39.prompt(w,c,Attention,TrellisUnfusedMLP,Reduce,Scatter,Plain),
|
|
858
|
+
d40.prompt(w,c,Delta,TrellisUnfusedMLP,Reduce,Scatter,Plain,qk),d41.prompt(w,c,Delta,TrellisUnfusedMLP,Reduce,Scatter,Plain,qk),d42.prompt(w,c,Delta,TrellisUnfusedMLP,Reduce,Scatter,Plain,qk),a43.prompt(w,c,Attention,TrellisUnfusedMLP,Reduce,Scatter,Plain),
|
|
859
|
+
d44.prompt(w,c,Delta,TrellisUnfusedMLP,Reduce,Scatter,Plain,qk),d45.prompt(w,c,Delta,TrellisUnfusedMLP,Reduce,Scatter,Plain,qk),d46.prompt(w,c,Delta,TrellisUnfusedMLP,Reduce,Scatter,Plain,qk),a47.prompt(w,c,Attention,TrellisUnfusedMLP,Reduce,Scatter,Plain),
|
|
860
|
+
d48.prompt(w,c,Delta,TrellisUnfusedMLP,Reduce,Scatter,Plain,qk),d49.prompt(w,c,Delta,TrellisUnfusedMLP,Reduce,Scatter,Plain,qk),d50.prompt(w,c,Delta,TrellisUnfusedMLP,Reduce,Scatter,Plain,qk),a51.prompt(w,c,Attention,TrellisUnfusedMLP,Reduce,Scatter,Plain),
|
|
861
|
+
d52.prompt(w,c,Delta,TrellisUnfusedMLP,Reduce,Scatter,Plain,qk),d53.prompt(w,c,Delta,TrellisUnfusedMLP,Reduce,Scatter,Plain,qk),d54.prompt(w,c,Delta,TrellisUnfusedMLP,Reduce,Scatter,Plain,qk),a55.prompt(w,c,Attention,TrellisUnfusedMLP,Reduce,Scatter,Plain),
|
|
862
|
+
d56.prompt(w,c,Delta,TrellisUnfusedMLP,Reduce,Scatter,Plain,qk),d57.prompt(w,c,Delta,TrellisUnfusedMLP,Reduce,Scatter,Plain,qk),d58.prompt(w,c,Delta,TrellisUnfusedMLP,Reduce,Scatter,Plain,qk),a59.prompt(w,c,Attention,TrellisUnfusedMLP,Reduce,Scatter,Plain),
|
|
863
|
+
d60.prompt(w,c,Delta,TrellisUnfusedMLP,Reduce,Scatter,Plain,qk),d61.prompt(w,c,Delta,TrellisUnfusedMLP,Reduce,Scatter,Plain,qk),d62.prompt(w,c,Delta,TrellisUnfusedMLP,Reduce,Scatter,Plain,qk),a63.prompt(w,c,Attention,TrellisUnfusedMLP,Reduce,Scatter,Plain),
|
|
864
|
+
];
|
|
865
|
+
const wide=this.#wide,mma=this.#mma;
|
|
866
|
+
const prompt512:Layers=[
|
|
867
|
+
d0.withProjections(w,c,Delta,expanded[0].mlp,Dense,Dense,Plain,Plain,Dense,qk),d1.withProjections(w,c,Delta,expanded[1].mlp,Dense,Dense,Plain,Plain,Dense,qk),d2.withProjections(w,c,Delta,expanded[2].mlp,Dense,Dense,Plain,Plain,Dense,qk),a3.withProjections(w,c,Attention,expanded[3].mlp,Dense,Plain,Plain,Dense),
|
|
868
|
+
d4.withProjections(w,c,Delta,expanded[4].mlp,Dense,Dense,Plain,Plain,Dense,qk),d5.withProjections(w,c,Delta,expanded[5].mlp,Dense,Dense,Plain,Plain,Dense,qk),d6.withProjections(w,c,Delta,expanded[6].mlp,Dense,Dense,Plain,Plain,Dense,qk),a7.withProjections(w,c,Attention,expanded[7].mlp,Dense,Plain,Plain,Dense),
|
|
869
|
+
d8.withProjections(w,c,Delta,expanded[8].mlp,Dense,Dense,Plain,Plain,Dense,qk),d9.withProjections(w,c,Delta,expanded[9].mlp,Dense,Dense,Plain,Plain,Dense,qk),d10.withProjections(w,c,Delta,expanded[10].mlp,Dense,Dense,Plain,Plain,Dense,qk),a11.withProjections(w,c,Attention,expanded[11].mlp,Dense,Plain,Plain,Dense),
|
|
870
|
+
d12.withProjections(w,c,Delta,expanded[12].mlp,Dense,Dense,Plain,Plain,Dense,qk),d13.withProjections(w,c,Delta,expanded[13].mlp,Dense,Dense,Plain,Plain,Dense,qk),d14.withProjections(w,c,Delta,expanded[14].mlp,Dense,Dense,Plain,Plain,Dense,qk),a15.withProjections(w,c,Attention,expanded[15].mlp,Dense,Plain,Plain,Dense),
|
|
871
|
+
d16.withProjections(w,c,Delta,expanded[16].mlp,Dense,Dense,Plain,Plain,Dense,qk),d17.withProjections(w,c,Delta,expanded[17].mlp,Dense,Dense,Plain,Plain,Dense,qk),d18.withProjections(w,c,Delta,expanded[18].mlp,Dense,Dense,Plain,Plain,Dense,qk),a19.withProjections(w,c,Attention,expanded[19].mlp,Dense,Plain,Plain,Dense),
|
|
872
|
+
d20.withProjections(w,c,Delta,expanded[20].mlp,Dense,Dense,Plain,Plain,Dense,qk),d21.withProjections(w,c,Delta,expanded[21].mlp,Dense,Dense,Plain,Plain,Dense,qk),d22.withProjections(w,c,Delta,expanded[22].mlp,Dense,Dense,Plain,Plain,Dense,qk),a23.withProjections(w,c,Attention,expanded[23].mlp,Dense,Plain,Plain,Dense),
|
|
873
|
+
d24.withProjections(w,c,Delta,expanded[24].mlp,Dense,Dense,Plain,Plain,Dense,qk),d25.withProjections(w,c,Delta,expanded[25].mlp,Dense,Dense,Plain,Plain,Dense,qk),d26.withProjections(w,c,Delta,expanded[26].mlp,Dense,Dense,Plain,Plain,Dense,qk),a27.withProjections(w,c,Attention,expanded[27].mlp,Dense,Plain,Plain,Dense),
|
|
874
|
+
d28.withProjections(w,c,Delta,expanded[28].mlp,Dense,Dense,Plain,Plain,Dense,qk),d29.withProjections(w,c,Delta,expanded[29].mlp,Dense,Dense,Plain,Plain,Dense,qk),d30.withProjections(w,c,Delta,expanded[30].mlp,Dense,Dense,Plain,Plain,Dense,qk),a31.withProjections(w,c,Attention,expanded[31].mlp,Dense,Plain,Plain,Dense),
|
|
875
|
+
d32.withProjections(w,c,Delta,expanded[32].mlp,Dense,Dense,Plain,Plain,Dense,qk),d33.withProjections(w,c,Delta,expanded[33].mlp,Dense,Dense,Plain,Plain,Dense,qk),d34.withProjections(w,c,Delta,expanded[34].mlp,Dense,Dense,Plain,Plain,Dense,qk),a35.withProjections(w,c,Attention,expanded[35].mlp,Dense,Plain,Plain,Dense),
|
|
876
|
+
d36.withProjections(w,c,Delta,expanded[36].mlp,Dense,Dense,Plain,Plain,Dense,qk),d37.withProjections(w,c,Delta,expanded[37].mlp,Dense,Dense,Plain,Plain,Dense,qk),d38.withProjections(w,c,Delta,expanded[38].mlp,Dense,Dense,Plain,Plain,Dense,qk),a39.withProjections(w,c,Attention,expanded[39].mlp,Dense,Plain,Plain,Dense),
|
|
877
|
+
d40.withProjections(w,c,Delta,expanded[40].mlp,Dense,Dense,Plain,Plain,Dense,qk),d41.withProjections(w,c,Delta,expanded[41].mlp,Dense,Dense,Plain,Plain,Dense,qk),d42.withProjections(w,c,Delta,expanded[42].mlp,Dense,Dense,Plain,Plain,Dense,qk),a43.withProjections(w,c,Attention,expanded[43].mlp,Dense,Plain,Plain,Dense),
|
|
878
|
+
d44.withProjections(w,c,Delta,expanded[44].mlp,Dense,Dense,Plain,Plain,Dense,qk),d45.withProjections(w,c,Delta,expanded[45].mlp,Dense,Dense,Plain,Plain,Dense,qk),d46.withProjections(w,c,Delta,expanded[46].mlp,Dense,Dense,Plain,Plain,Dense,qk),a47.withProjections(w,c,Attention,expanded[47].mlp,Dense,Plain,Plain,Dense),
|
|
879
|
+
d48.withProjections(w,c,Delta,expanded[48].mlp,Dense,Dense,Plain,Plain,Dense,qk),d49.withProjections(w,c,Delta,expanded[49].mlp,Dense,Dense,Plain,Plain,Dense,qk),d50.withProjections(w,c,Delta,expanded[50].mlp,Dense,Dense,Plain,Plain,Dense,qk),a51.withProjections(w,c,Attention,expanded[51].mlp,Dense,Plain,Plain,Dense),
|
|
880
|
+
d52.withProjections(w,c,Delta,expanded[52].mlp,Dense,Dense,Plain,Plain,Dense,qk),d53.withProjections(w,c,Delta,expanded[53].mlp,Dense,Dense,Plain,Plain,Dense,qk),d54.withProjections(w,c,Delta,expanded[54].mlp,Dense,Dense,Plain,Plain,Dense,qk),a55.withProjections(w,c,Attention,expanded[55].mlp,Dense,Plain,Plain,Dense),
|
|
881
|
+
d56.withProjections(w,c,Delta,expanded[56].mlp,Dense,Dense,Plain,Plain,Dense,qk),d57.withProjections(w,c,Delta,expanded[57].mlp,Dense,Dense,Plain,Plain,Dense,qk),d58.withProjections(w,c,Delta,expanded[58].mlp,Dense,Dense,Plain,Plain,Dense,qk),a59.withProjections(w,c,Attention,expanded[59].mlp,Dense,Plain,Plain,Dense),
|
|
882
|
+
d60.withProjections(w,c,Delta,expanded[60].mlp,Dense,Dense,Plain,Plain,Dense,qk),d61.withProjections(w,c,Delta,expanded[61].mlp,Dense,Dense,Plain,Plain,Dense,qk),d62.withProjections(w,c,Delta,expanded[62].mlp,Dense,Dense,Plain,Plain,Dense,qk),a63.withProjections(w,c,Attention,expanded[63].mlp,Dense,Plain,Plain,Dense),
|
|
883
|
+
];
|
|
884
|
+
const promptDense:Layers=[
|
|
885
|
+
prompt512[0],prompt512[1],prompt512[2],a3.withProjections(w,c,Attention,expanded[3].mlp,Dense,Dense,Dense,Dense),
|
|
886
|
+
prompt512[4],prompt512[5],prompt512[6],a7.withProjections(w,c,Attention,expanded[7].mlp,Dense,Dense,Dense,Dense),
|
|
887
|
+
prompt512[8],prompt512[9],prompt512[10],a11.withProjections(w,c,Attention,expanded[11].mlp,Dense,Dense,Dense,Dense),
|
|
888
|
+
prompt512[12],prompt512[13],prompt512[14],a15.withProjections(w,c,Attention,expanded[15].mlp,Dense,Dense,Dense,Dense),
|
|
889
|
+
prompt512[16],prompt512[17],prompt512[18],a19.withProjections(w,c,Attention,expanded[19].mlp,Dense,Dense,Dense,Dense),
|
|
890
|
+
prompt512[20],prompt512[21],prompt512[22],a23.withProjections(w,c,Attention,expanded[23].mlp,Dense,Dense,Dense,Dense),
|
|
891
|
+
prompt512[24],prompt512[25],prompt512[26],a27.withProjections(w,c,Attention,expanded[27].mlp,Dense,Dense,Dense,Dense),
|
|
892
|
+
prompt512[28],prompt512[29],prompt512[30],a31.withProjections(w,c,Attention,expanded[31].mlp,Dense,Dense,Dense,Dense),
|
|
893
|
+
prompt512[32],prompt512[33],prompt512[34],a35.withProjections(w,c,Attention,expanded[35].mlp,Dense,Dense,Dense,Dense),
|
|
894
|
+
prompt512[36],prompt512[37],prompt512[38],a39.withProjections(w,c,Attention,expanded[39].mlp,Dense,Dense,Dense,Dense),
|
|
895
|
+
prompt512[40],prompt512[41],prompt512[42],a43.withProjections(w,c,Attention,expanded[43].mlp,Dense,Dense,Dense,Dense),
|
|
896
|
+
prompt512[44],prompt512[45],prompt512[46],a47.withProjections(w,c,Attention,expanded[47].mlp,Dense,Dense,Dense,Dense),
|
|
897
|
+
prompt512[48],prompt512[49],prompt512[50],a51.withProjections(w,c,Attention,expanded[51].mlp,Dense,Dense,Dense,Dense),
|
|
898
|
+
prompt512[52],prompt512[53],prompt512[54],a55.withProjections(w,c,Attention,expanded[55].mlp,Dense,Dense,Dense,Dense),
|
|
899
|
+
prompt512[56],prompt512[57],prompt512[58],a59.withProjections(w,c,Attention,expanded[59].mlp,Dense,Dense,Dense,Dense),
|
|
900
|
+
prompt512[60],prompt512[61],prompt512[62],a63.withProjections(w,c,Attention,expanded[63].mlp,Dense,Dense,Dense,Dense),
|
|
901
|
+
];
|
|
902
|
+
const mixedWide:Layers=[
|
|
903
|
+
d0.withMlp(w,c,Delta,wide[0].mlp,Plain,qk),d1.withMlp(w,c,Delta,wide[1].mlp,Plain,qk),d2.withMlp(w,c,Delta,wide[2].mlp,Plain,qk),a3.withMlp(w,c,Attention,wide[3].mlp,Plain),
|
|
904
|
+
d4.withMlp(w,c,Delta,wide[4].mlp,Plain,qk),d5.withMlp(w,c,Delta,wide[5].mlp,Plain,qk),d6.withMlp(w,c,Delta,wide[6].mlp,Plain,qk),a7.withMlp(w,c,Attention,wide[7].mlp,Plain),
|
|
905
|
+
d8.withMlp(w,c,Delta,wide[8].mlp,Plain,qk),d9.withMlp(w,c,Delta,wide[9].mlp,Plain,qk),d10.withMlp(w,c,Delta,wide[10].mlp,Plain,qk),a11.withMlp(w,c,Attention,wide[11].mlp,Plain),
|
|
906
|
+
d12.withMlp(w,c,Delta,wide[12].mlp,Plain,qk),d13.withMlp(w,c,Delta,wide[13].mlp,Plain,qk),d14.withMlp(w,c,Delta,wide[14].mlp,Plain,qk),a15.withMlp(w,c,Attention,wide[15].mlp,Plain),
|
|
907
|
+
d16.withMlp(w,c,Delta,wide[16].mlp,Plain,qk),d17.withMlp(w,c,Delta,wide[17].mlp,Plain,qk),d18.withMlp(w,c,Delta,wide[18].mlp,Plain,qk),a19.withMlp(w,c,Attention,wide[19].mlp,Plain),
|
|
908
|
+
d20.withMlp(w,c,Delta,wide[20].mlp,Plain,qk),d21.withMlp(w,c,Delta,wide[21].mlp,Plain,qk),d22.withMlp(w,c,Delta,wide[22].mlp,Plain,qk),a23.withMlp(w,c,Attention,wide[23].mlp,Plain),
|
|
909
|
+
d24.withMlp(w,c,Delta,wide[24].mlp,Plain,qk),d25.withMlp(w,c,Delta,wide[25].mlp,Plain,qk),d26.withMlp(w,c,Delta,wide[26].mlp,Plain,qk),a27.withMlp(w,c,Attention,wide[27].mlp,Plain),
|
|
910
|
+
d28.withMlp(w,c,Delta,wide[28].mlp,Plain,qk),d29.withMlp(w,c,Delta,wide[29].mlp,Plain,qk),d30.withMlp(w,c,Delta,wide[30].mlp,Plain,qk),a31.withMlp(w,c,Attention,wide[31].mlp,Plain),
|
|
911
|
+
d32.withMlp(w,c,Delta,wide[32].mlp,Plain,qk),d33.withMlp(w,c,Delta,wide[33].mlp,Plain,qk),d34.withMlp(w,c,Delta,wide[34].mlp,Plain,qk),a35.withMlp(w,c,Attention,wide[35].mlp,Plain),
|
|
912
|
+
d36.withMlp(w,c,Delta,wide[36].mlp,Plain,qk),d37.withMlp(w,c,Delta,wide[37].mlp,Plain,qk),d38.withMlp(w,c,Delta,wide[38].mlp,Plain,qk),a39.withMlp(w,c,Attention,wide[39].mlp,Plain),
|
|
913
|
+
d40.withMlp(w,c,Delta,wide[40].mlp,Plain,qk),d41.withMlp(w,c,Delta,wide[41].mlp,Plain,qk),d42.withMlp(w,c,Delta,wide[42].mlp,Plain,qk),a43.withMlp(w,c,Attention,wide[43].mlp,Plain),
|
|
914
|
+
d44.withMlp(w,c,Delta,wide[44].mlp,Plain,qk),d45.withMlp(w,c,Delta,wide[45].mlp,Plain,qk),d46.withMlp(w,c,Delta,wide[46].mlp,Plain,qk),a47.withMlp(w,c,Attention,wide[47].mlp,Plain),
|
|
915
|
+
d48.withMlp(w,c,Delta,wide[48].mlp,Plain,qk),d49.withMlp(w,c,Delta,wide[49].mlp,Plain,qk),d50.withMlp(w,c,Delta,wide[50].mlp,Plain,qk),a51.withMlp(w,c,Attention,wide[51].mlp,Plain),
|
|
916
|
+
d52.withMlp(w,c,Delta,wide[52].mlp,Plain,qk),d53.withMlp(w,c,Delta,wide[53].mlp,Plain,qk),d54.withMlp(w,c,Delta,wide[54].mlp,Plain,qk),a55.withMlp(w,c,Attention,wide[55].mlp,Plain),
|
|
917
|
+
d56.withMlp(w,c,Delta,wide[56].mlp,Plain,qk),d57.withMlp(w,c,Delta,wide[57].mlp,Plain,qk),d58.withMlp(w,c,Delta,wide[58].mlp,Plain,qk),a59.withMlp(w,c,Attention,wide[59].mlp,Plain),
|
|
918
|
+
d60.withMlp(w,c,Delta,wide[60].mlp,Plain,qk),d61.withMlp(w,c,Delta,wide[61].mlp,Plain,qk),d62.withMlp(w,c,Delta,wide[62].mlp,Plain,qk),a63.withMlp(w,c,Attention,wide[63].mlp,Plain),
|
|
919
|
+
];
|
|
920
|
+
const mediaStock:Layers=[
|
|
921
|
+
stock[0],stock[1],stock[2],a3.mediaLayer(w,c,Attention,stock[3].mlp,Plain),
|
|
922
|
+
stock[4],stock[5],stock[6],a7.mediaLayer(w,c,Attention,stock[7].mlp,Plain),
|
|
923
|
+
stock[8],stock[9],stock[10],a11.mediaLayer(w,c,Attention,stock[11].mlp,Plain),
|
|
924
|
+
stock[12],stock[13],stock[14],a15.mediaLayer(w,c,Attention,stock[15].mlp,Plain),
|
|
925
|
+
stock[16],stock[17],stock[18],a19.mediaLayer(w,c,Attention,stock[19].mlp,Plain),
|
|
926
|
+
stock[20],stock[21],stock[22],a23.mediaLayer(w,c,Attention,stock[23].mlp,Plain),
|
|
927
|
+
stock[24],stock[25],stock[26],a27.mediaLayer(w,c,Attention,stock[27].mlp,Plain),
|
|
928
|
+
stock[28],stock[29],stock[30],a31.mediaLayer(w,c,Attention,stock[31].mlp,Plain),
|
|
929
|
+
stock[32],stock[33],stock[34],a35.mediaLayer(w,c,Attention,stock[35].mlp,Plain),
|
|
930
|
+
stock[36],stock[37],stock[38],a39.mediaLayer(w,c,Attention,stock[39].mlp,Plain),
|
|
931
|
+
stock[40],stock[41],stock[42],a43.mediaLayer(w,c,Attention,stock[43].mlp,Plain),
|
|
932
|
+
stock[44],stock[45],stock[46],a47.mediaLayer(w,c,Attention,stock[47].mlp,Plain),
|
|
933
|
+
stock[48],stock[49],stock[50],a51.mediaLayer(w,c,Attention,stock[51].mlp,Plain),
|
|
934
|
+
stock[52],stock[53],stock[54],a55.mediaLayer(w,c,Attention,stock[55].mlp,Plain),
|
|
935
|
+
stock[56],stock[57],stock[58],a59.mediaLayer(w,c,Attention,stock[59].mlp,Plain),
|
|
936
|
+
stock[60],stock[61],stock[62],a63.mediaLayer(w,c,Attention,stock[63].mlp,Plain),
|
|
937
|
+
];
|
|
938
|
+
const mediaFive:Layers=[
|
|
939
|
+
five[0],five[1],five[2],a3.mediaLayer(w,c,Attention,five[3].mlp,Plain),
|
|
940
|
+
five[4],five[5],five[6],a7.mediaLayer(w,c,Attention,five[7].mlp,Plain),
|
|
941
|
+
five[8],five[9],five[10],a11.mediaLayer(w,c,Attention,five[11].mlp,Plain),
|
|
942
|
+
five[12],five[13],five[14],a15.mediaLayer(w,c,Attention,five[15].mlp,Plain),
|
|
943
|
+
five[16],five[17],five[18],a19.mediaLayer(w,c,Attention,five[19].mlp,Plain),
|
|
944
|
+
five[20],five[21],five[22],a23.mediaLayer(w,c,Attention,five[23].mlp,Plain),
|
|
945
|
+
five[24],five[25],five[26],a27.mediaLayer(w,c,Attention,five[27].mlp,Plain),
|
|
946
|
+
five[28],five[29],five[30],a31.mediaLayer(w,c,Attention,five[31].mlp,Plain),
|
|
947
|
+
five[32],five[33],five[34],a35.mediaLayer(w,c,Attention,five[35].mlp,Plain),
|
|
948
|
+
five[36],five[37],five[38],a39.mediaLayer(w,c,Attention,five[39].mlp,Plain),
|
|
949
|
+
five[40],five[41],five[42],a43.mediaLayer(w,c,Attention,five[43].mlp,Plain),
|
|
950
|
+
five[44],five[45],five[46],a47.mediaLayer(w,c,Attention,five[47].mlp,Plain),
|
|
951
|
+
five[48],five[49],five[50],a51.mediaLayer(w,c,Attention,five[51].mlp,Plain),
|
|
952
|
+
five[52],five[53],five[54],a55.mediaLayer(w,c,Attention,five[55].mlp,Plain),
|
|
953
|
+
five[56],five[57],five[58],a59.mediaLayer(w,c,Attention,five[59].mlp,Plain),
|
|
954
|
+
five[60],five[61],five[62],a63.mediaLayer(w,c,Attention,five[63].mlp,Plain),
|
|
955
|
+
];
|
|
956
|
+
const mediaTiled:Layers=[
|
|
957
|
+
tiled[0],tiled[1],tiled[2],a3.mediaLayer(w,c,Attention,tiled[3].mlp,Plain),
|
|
958
|
+
tiled[4],tiled[5],tiled[6],a7.mediaLayer(w,c,Attention,tiled[7].mlp,Plain),
|
|
959
|
+
tiled[8],tiled[9],tiled[10],a11.mediaLayer(w,c,Attention,tiled[11].mlp,Plain),
|
|
960
|
+
tiled[12],tiled[13],tiled[14],a15.mediaLayer(w,c,Attention,tiled[15].mlp,Plain),
|
|
961
|
+
tiled[16],tiled[17],tiled[18],a19.mediaLayer(w,c,Attention,tiled[19].mlp,Plain),
|
|
962
|
+
tiled[20],tiled[21],tiled[22],a23.mediaLayer(w,c,Attention,tiled[23].mlp,Plain),
|
|
963
|
+
tiled[24],tiled[25],tiled[26],a27.mediaLayer(w,c,Attention,tiled[27].mlp,Plain),
|
|
964
|
+
tiled[28],tiled[29],tiled[30],a31.mediaLayer(w,c,Attention,tiled[31].mlp,Plain),
|
|
965
|
+
tiled[32],tiled[33],tiled[34],a35.mediaLayer(w,c,Attention,tiled[35].mlp,Plain),
|
|
966
|
+
tiled[36],tiled[37],tiled[38],a39.mediaLayer(w,c,Attention,tiled[39].mlp,Plain),
|
|
967
|
+
tiled[40],tiled[41],tiled[42],a43.mediaLayer(w,c,Attention,tiled[43].mlp,Plain),
|
|
968
|
+
tiled[44],tiled[45],tiled[46],a47.mediaLayer(w,c,Attention,tiled[47].mlp,Plain),
|
|
969
|
+
tiled[48],tiled[49],tiled[50],a51.mediaLayer(w,c,Attention,tiled[51].mlp,Plain),
|
|
970
|
+
tiled[52],tiled[53],tiled[54],a55.mediaLayer(w,c,Attention,tiled[55].mlp,Plain),
|
|
971
|
+
tiled[56],tiled[57],tiled[58],a59.mediaLayer(w,c,Attention,tiled[59].mlp,Plain),
|
|
972
|
+
tiled[60],tiled[61],tiled[62],a63.mediaLayer(w,c,Attention,tiled[63].mlp,Plain),
|
|
973
|
+
];
|
|
974
|
+
const mediaExpanded:Layers=[
|
|
975
|
+
expanded[0],expanded[1],expanded[2],a3.mediaLayer(w,c,Attention,expanded[3].mlp,Plain),
|
|
976
|
+
expanded[4],expanded[5],expanded[6],a7.mediaLayer(w,c,Attention,expanded[7].mlp,Plain),
|
|
977
|
+
expanded[8],expanded[9],expanded[10],a11.mediaLayer(w,c,Attention,expanded[11].mlp,Plain),
|
|
978
|
+
expanded[12],expanded[13],expanded[14],a15.mediaLayer(w,c,Attention,expanded[15].mlp,Plain),
|
|
979
|
+
expanded[16],expanded[17],expanded[18],a19.mediaLayer(w,c,Attention,expanded[19].mlp,Plain),
|
|
980
|
+
expanded[20],expanded[21],expanded[22],a23.mediaLayer(w,c,Attention,expanded[23].mlp,Plain),
|
|
981
|
+
expanded[24],expanded[25],expanded[26],a27.mediaLayer(w,c,Attention,expanded[27].mlp,Plain),
|
|
982
|
+
expanded[28],expanded[29],expanded[30],a31.mediaLayer(w,c,Attention,expanded[31].mlp,Plain),
|
|
983
|
+
expanded[32],expanded[33],expanded[34],a35.mediaLayer(w,c,Attention,expanded[35].mlp,Plain),
|
|
984
|
+
expanded[36],expanded[37],expanded[38],a39.mediaLayer(w,c,Attention,expanded[39].mlp,Plain),
|
|
985
|
+
expanded[40],expanded[41],expanded[42],a43.mediaLayer(w,c,Attention,expanded[43].mlp,Plain),
|
|
986
|
+
expanded[44],expanded[45],expanded[46],a47.mediaLayer(w,c,Attention,expanded[47].mlp,Plain),
|
|
987
|
+
expanded[48],expanded[49],expanded[50],a51.mediaLayer(w,c,Attention,expanded[51].mlp,Plain),
|
|
988
|
+
expanded[52],expanded[53],expanded[54],a55.mediaLayer(w,c,Attention,expanded[55].mlp,Plain),
|
|
989
|
+
expanded[56],expanded[57],expanded[58],a59.mediaLayer(w,c,Attention,expanded[59].mlp,Plain),
|
|
990
|
+
expanded[60],expanded[61],expanded[62],a63.mediaLayer(w,c,Attention,expanded[63].mlp,Plain),
|
|
991
|
+
];
|
|
992
|
+
const mmaFive=this.#mmaFive;
|
|
993
|
+
const mixedWideNarrow:Layers=[
|
|
994
|
+
d0.stage(w,c,Delta,GateDownMLP,WideGate,Reduce,WideDown,Plain,Plain,Plain,Plain,Plain,qk),d1.stage(w,c,Delta,GateDownMLP,WideGate,Reduce,WideDownK3i,Plain,Plain,Plain,Plain,Plain,qk),d2.stage(w,c,Delta,GateDownMLP,WideGate,Reduce,WideDownK3i,Plain,Plain,Plain,Plain,Plain,qk),a3.stage(w,c,Attention,GateDownMLP,WideGate,Reduce,WideDownK3i,Plain,Plain,Plain,Plain),
|
|
995
|
+
d4.stage(w,c,Delta,GateDownMLP,WideGate,Reduce,WideDownK3i,Plain,Plain,Plain,Plain,Plain,qk),d5.stage(w,c,Delta,GateDownMLP,WideGate,Reduce,WideDownK3i,Plain,Plain,Plain,Plain,Plain,qk),d6.stage(w,c,Delta,GateDownMLP,WideGate,Reduce,WideDownK3i,Plain,Plain,Plain,Plain,Plain,qk),a7.stage(w,c,Attention,GateDownMLP,WideGate,Reduce,WideDownK3i,Plain,Plain,Plain,Plain),
|
|
996
|
+
d8.stage(w,c,Delta,GateDownMLP,WideGate,Reduce,WideDownK3i,Plain,Plain,Plain,Plain,Plain,qk),d9.stage(w,c,Delta,GateDownMLP,WideGate,Reduce,WideDownK3i,Plain,Plain,Plain,Plain,Plain,qk),d10.stage(w,c,Delta,GateDownMLP,WideGate,Reduce,WideDownK3i,Plain,Plain,Plain,Plain,Plain,qk),a11.stage(w,c,Attention,GateDownMLP,WideGate,Reduce,WideDown,Plain,Plain,Plain,Plain),
|
|
997
|
+
d12.stage(w,c,Delta,GateDownMLP,WideGate,Reduce,WideDownK3i,Plain,Plain,Plain,Plain,Plain,qk),d13.stage(w,c,Delta,GateDownMLP,WideGate,Reduce,WideDown,Plain,Plain,Plain,Plain,Plain,qk),d14.stage(w,c,Delta,GateDownMLP,WideGate,Reduce,WideDownK3i,Plain,Plain,Plain,Plain,Plain,qk),a15.stage(w,c,Attention,GateDownMLP,WideGate,Reduce,WideDown,Plain,Plain,Plain,Plain),
|
|
998
|
+
d16.stage(w,c,Delta,GateDownMLP,WideGate,Reduce,WideDown,Plain,Plain,Plain,Plain,Plain,qk),d17.stage(w,c,Delta,GateDownMLP,WideGate,Reduce,WideDown,Plain,Plain,Plain,Plain,Plain,qk),d18.stage(w,c,Delta,GateDownMLP,WideGate,Reduce,WideDown,Plain,Plain,Plain,Plain,Plain,qk),a19.stage(w,c,Attention,GateDownMLP,WideGate,Reduce,WideDownK3i,Plain,Plain,Plain,Plain),
|
|
999
|
+
d20.stage(w,c,Delta,GateDownMLP,WideGate,Reduce,WideDownK3i,Plain,Plain,Plain,Plain,Plain,qk),d21.stage(w,c,Delta,GateDownMLP,WideGate,Reduce,WideDownK3i,Plain,Plain,Plain,Plain,Plain,qk),d22.stage(w,c,Delta,GateDownMLP,WideGate,Reduce,WideDownK3i,Plain,Plain,Plain,Plain,Plain,qk),a23.stage(w,c,Attention,GateDownMLP,WideGate,Reduce,WideDownK3i,Plain,Plain,Plain,Plain),
|
|
1000
|
+
d24.stage(w,c,Delta,GateDownMLP,WideGate,Reduce,WideDownK3i,Plain,Plain,Plain,Plain,Plain,qk),d25.stage(w,c,Delta,GateDownMLP,WideGate,Reduce,WideDownK3i,Plain,Plain,Plain,Plain,Plain,qk),d26.stage(w,c,Delta,GateDownMLP,WideGate,Reduce,WideDownK3i,Plain,Plain,Plain,Plain,Plain,qk),a27.stage(w,c,Attention,GateDownMLP,WideGate,Reduce,WideDownK3i,Plain,Plain,Plain,Plain),
|
|
1001
|
+
d28.stage(w,c,Delta,GateDownMLP,WideGate,Reduce,WideDownK3i,Plain,Plain,Plain,Plain,Plain,qk),d29.stage(w,c,Delta,GateDownMLP,WideGate,Reduce,WideDown,Plain,Plain,Plain,Plain,Plain,qk),d30.stage(w,c,Delta,GateDownMLP,WideGate,Reduce,WideDownK3i,Plain,Plain,Plain,Plain,Plain,qk),a31.stage(w,c,Attention,GateDownMLP,WideGate,Reduce,WideDown,Plain,Plain,Plain,Plain),
|
|
1002
|
+
d32.stage(w,c,Delta,GateDownMLP,WideGate,Reduce,WideDown,Plain,Plain,Plain,Plain,Plain,qk),d33.stage(w,c,Delta,GateDownMLP,WideGate,Reduce,WideDown,Plain,Plain,Plain,Plain,Plain,qk),d34.stage(w,c,Delta,GateDownMLP,WideGate,Reduce,WideDownK3i,Plain,Plain,Plain,Plain,Plain,qk),a35.stage(w,c,Attention,GateDownMLP,WideGate,Reduce,WideDownK3i,Plain,Plain,Plain,Plain),
|
|
1003
|
+
d36.stage(w,c,Delta,GateDownMLP,WideGate,Reduce,WideDownK3i,Plain,Plain,Plain,Plain,Plain,qk),d37.stage(w,c,Delta,GateDownMLP,WideGate,Reduce,WideDownK3i,Plain,Plain,Plain,Plain,Plain,qk),d38.stage(w,c,Delta,GateDownMLP,WideGate,Reduce,WideDownK3i,Plain,Plain,Plain,Plain,Plain,qk),a39.stage(w,c,Attention,GateDownMLP,WideGate,Reduce,WideDownK3i,Plain,Plain,Plain,Plain),
|
|
1004
|
+
d40.stage(w,c,Delta,GateDownMLP,WideGate,Reduce,WideDownK3i,Plain,Plain,Plain,Plain,Plain,qk),d41.stage(w,c,Delta,GateDownMLP,WideGate,Reduce,WideDownK3i,Plain,Plain,Plain,Plain,Plain,qk),d42.stage(w,c,Delta,GateDownMLP,WideGate,Reduce,WideDownK3i,Plain,Plain,Plain,Plain,Plain,qk),a43.stage(w,c,Attention,GateDownMLP,WideGate,Reduce,WideDownK3i,Plain,Plain,Plain,Plain),
|
|
1005
|
+
d44.stage(w,c,Delta,GateDownMLP,WideGate,Reduce,WideDownK3i,Plain,Plain,Plain,Plain,Plain,qk),d45.stage(w,c,Delta,GateDownMLP,WideGate,Reduce,WideDown,Plain,Plain,Plain,Plain,Plain,qk),d46.stage(w,c,Delta,GateDownMLP,WideGate,Reduce,WideDown,Plain,Plain,Plain,Plain,Plain,qk),a47.stage(w,c,Attention,GateDownMLP,WideGate,Reduce,WideDown,Plain,Plain,Plain,Plain),
|
|
1006
|
+
d48.stage(w,c,Delta,GateDownMLP,WideGate,Reduce,WideDown,Plain,Plain,Plain,Plain,Plain,qk),d49.stage(w,c,Delta,GateDownMLP,WideGate,Reduce,WideDown,Plain,Plain,Plain,Plain,Plain,qk),d50.stage(w,c,Delta,GateDownMLP,WideGate,Reduce,WideDownK3i,Plain,Plain,Plain,Plain,Plain,qk),a51.stage(w,c,Attention,GateDownMLP,WideGate,Reduce,WideDownK3i,Plain,Plain,Plain,Plain),
|
|
1007
|
+
d52.stage(w,c,Delta,GateDownMLP,WideGate,Reduce,WideDownK3i,Plain,Plain,Plain,Plain,Plain,qk),d53.stage(w,c,Delta,GateDownMLP,WideGate,Reduce,WideDownK3i,Plain,Plain,Plain,Plain,Plain,qk),d54.stage(w,c,Delta,GateDownMLP,WideGate,Reduce,WideDown,Plain,Plain,Plain,Plain,Plain,qk),a55.stage(w,c,Attention,GateDownMLP,WideGate,Reduce,WideDown,Plain,Plain,Plain,Plain),
|
|
1008
|
+
d56.stage(w,c,Delta,GateDownMLP,WideGate,Reduce,WideDown,Plain,Plain,Plain,Plain,Plain,qk),d57.stage(w,c,Delta,GateDownMLP,WideGate,Reduce,WideDown,Plain,Plain,Plain,Plain,Plain,qk),d58.stage(w,c,Delta,GateDownMLP,WideGate,Reduce,WideDown,Plain,Plain,Plain,Plain,Plain,qk),a59.stage(w,c,Attention,GateDownMLP,WideGate,Reduce,WideDown,Plain,Plain,Plain,Plain),
|
|
1009
|
+
d60.stage(w,c,Delta,GateDownMLP,WideGate,Reduce,WideDown,Plain,Plain,Plain,Plain,Plain,qk),d61.stage(w,c,Delta,GateDownMLP,WideGate,Reduce,WideDown,Plain,Plain,Plain,Plain,Plain,qk),d62.stage(w,c,Delta,GateDownMLP,WideGate,Reduce,WideDownK3i,Plain,Plain,Plain,Plain,Plain,qk),a63.stage(w,c,Attention,GateDownMLP,WideGate,Reduce,WideDownK3i,Plain,Plain,Plain,Plain),
|
|
1010
|
+
];
|
|
1011
|
+
const mixedWideFive:Layers=[
|
|
1012
|
+
d0.stage(w,c,Delta,GateDownMLP,WideGate,Tiled,WideDown,Plain,Plain,Plain,Plain,Plain,qk),d1.stage(w,c,Delta,GateDownMLP,WideGate,Tiled,WideDownK3i,Plain,Plain,Plain,Plain,Plain,qk),d2.stage(w,c,Delta,GateDownMLP,WideGate,Tiled,WideDownK3i,Plain,Plain,Plain,Plain,Plain,qk),a3.stage(w,c,Attention,GateDownMLP,WideGate,Tiled,WideDownK3i,Plain,Plain,Plain,Plain),
|
|
1013
|
+
d4.stage(w,c,Delta,GateDownMLP,WideGate,Tiled,WideDownK3i,Plain,Plain,Plain,Plain,Plain,qk),d5.stage(w,c,Delta,GateDownMLP,WideGate,Tiled,WideDownK3i,Plain,Plain,Plain,Plain,Plain,qk),d6.stage(w,c,Delta,GateDownMLP,WideGate,Tiled,WideDownK3i,Plain,Plain,Plain,Plain,Plain,qk),a7.stage(w,c,Attention,GateDownMLP,WideGate,Tiled,WideDownK3i,Plain,Plain,Plain,Plain),
|
|
1014
|
+
d8.stage(w,c,Delta,GateDownMLP,WideGate,Tiled,WideDownK3i,Plain,Plain,Plain,Plain,Plain,qk),d9.stage(w,c,Delta,GateDownMLP,WideGate,Tiled,WideDownK3i,Plain,Plain,Plain,Plain,Plain,qk),d10.stage(w,c,Delta,GateDownMLP,WideGate,Tiled,WideDownK3i,Plain,Plain,Plain,Plain,Plain,qk),a11.stage(w,c,Attention,GateDownMLP,WideGate,Tiled,WideDown,Plain,Plain,Plain,Plain),
|
|
1015
|
+
d12.stage(w,c,Delta,GateDownMLP,WideGate,Tiled,WideDownK3i,Plain,Plain,Plain,Plain,Plain,qk),d13.stage(w,c,Delta,GateDownMLP,WideGate,Tiled,WideDown,Plain,Plain,Plain,Plain,Plain,qk),d14.stage(w,c,Delta,GateDownMLP,WideGate,Tiled,WideDownK3i,Plain,Plain,Plain,Plain,Plain,qk),a15.stage(w,c,Attention,GateDownMLP,WideGate,Tiled,WideDown,Plain,Plain,Plain,Plain),
|
|
1016
|
+
d16.stage(w,c,Delta,GateDownMLP,WideGate,Tiled,WideDown,Plain,Plain,Plain,Plain,Plain,qk),d17.stage(w,c,Delta,GateDownMLP,WideGate,Tiled,WideDown,Plain,Plain,Plain,Plain,Plain,qk),d18.stage(w,c,Delta,GateDownMLP,WideGate,Tiled,WideDown,Plain,Plain,Plain,Plain,Plain,qk),a19.stage(w,c,Attention,GateDownMLP,WideGate,Tiled,WideDownK3i,Plain,Plain,Plain,Plain),
|
|
1017
|
+
d20.stage(w,c,Delta,GateDownMLP,WideGate,Tiled,WideDownK3i,Plain,Plain,Plain,Plain,Plain,qk),d21.stage(w,c,Delta,GateDownMLP,WideGate,Tiled,WideDownK3i,Plain,Plain,Plain,Plain,Plain,qk),d22.stage(w,c,Delta,GateDownMLP,WideGate,Tiled,WideDownK3i,Plain,Plain,Plain,Plain,Plain,qk),a23.stage(w,c,Attention,GateDownMLP,WideGate,Tiled,WideDownK3i,Plain,Plain,Plain,Plain),
|
|
1018
|
+
d24.stage(w,c,Delta,GateDownMLP,WideGate,Tiled,WideDownK3i,Plain,Plain,Plain,Plain,Plain,qk),d25.stage(w,c,Delta,GateDownMLP,WideGate,Tiled,WideDownK3i,Plain,Plain,Plain,Plain,Plain,qk),d26.stage(w,c,Delta,GateDownMLP,WideGate,Tiled,WideDownK3i,Plain,Plain,Plain,Plain,Plain,qk),a27.stage(w,c,Attention,GateDownMLP,WideGate,Tiled,WideDownK3i,Plain,Plain,Plain,Plain),
|
|
1019
|
+
d28.stage(w,c,Delta,GateDownMLP,WideGate,Tiled,WideDownK3i,Plain,Plain,Plain,Plain,Plain,qk),d29.stage(w,c,Delta,GateDownMLP,WideGate,Tiled,WideDown,Plain,Plain,Plain,Plain,Plain,qk),d30.stage(w,c,Delta,GateDownMLP,WideGate,Tiled,WideDownK3i,Plain,Plain,Plain,Plain,Plain,qk),a31.stage(w,c,Attention,GateDownMLP,WideGate,Tiled,WideDown,Plain,Plain,Plain,Plain),
|
|
1020
|
+
d32.stage(w,c,Delta,GateDownMLP,WideGate,Tiled,WideDown,Plain,Plain,Plain,Plain,Plain,qk),d33.stage(w,c,Delta,GateDownMLP,WideGate,Tiled,WideDown,Plain,Plain,Plain,Plain,Plain,qk),d34.stage(w,c,Delta,GateDownMLP,WideGate,Tiled,WideDownK3i,Plain,Plain,Plain,Plain,Plain,qk),a35.stage(w,c,Attention,GateDownMLP,WideGate,Tiled,WideDownK3i,Plain,Plain,Plain,Plain),
|
|
1021
|
+
d36.stage(w,c,Delta,GateDownMLP,WideGate,Tiled,WideDownK3i,Plain,Plain,Plain,Plain,Plain,qk),d37.stage(w,c,Delta,GateDownMLP,WideGate,Tiled,WideDownK3i,Plain,Plain,Plain,Plain,Plain,qk),d38.stage(w,c,Delta,GateDownMLP,WideGate,Tiled,WideDownK3i,Plain,Plain,Plain,Plain,Plain,qk),a39.stage(w,c,Attention,GateDownMLP,WideGate,Tiled,WideDownK3i,Plain,Plain,Plain,Plain),
|
|
1022
|
+
d40.stage(w,c,Delta,GateDownMLP,WideGate,Tiled,WideDownK3i,Plain,Plain,Plain,Plain,Plain,qk),d41.stage(w,c,Delta,GateDownMLP,WideGate,Tiled,WideDownK3i,Plain,Plain,Plain,Plain,Plain,qk),d42.stage(w,c,Delta,GateDownMLP,WideGate,Tiled,WideDownK3i,Plain,Plain,Plain,Plain,Plain,qk),a43.stage(w,c,Attention,GateDownMLP,WideGate,Tiled,WideDownK3i,Plain,Plain,Plain,Plain),
|
|
1023
|
+
d44.stage(w,c,Delta,GateDownMLP,WideGate,Tiled,WideDownK3i,Plain,Plain,Plain,Plain,Plain,qk),d45.stage(w,c,Delta,GateDownMLP,WideGate,Tiled,WideDown,Plain,Plain,Plain,Plain,Plain,qk),d46.stage(w,c,Delta,GateDownMLP,WideGate,Tiled,WideDown,Plain,Plain,Plain,Plain,Plain,qk),a47.stage(w,c,Attention,GateDownMLP,WideGate,Tiled,WideDown,Plain,Plain,Plain,Plain),
|
|
1024
|
+
d48.stage(w,c,Delta,GateDownMLP,WideGate,Tiled,WideDown,Plain,Plain,Plain,Plain,Plain,qk),d49.stage(w,c,Delta,GateDownMLP,WideGate,Tiled,WideDown,Plain,Plain,Plain,Plain,Plain,qk),d50.stage(w,c,Delta,GateDownMLP,WideGate,Tiled,WideDownK3i,Plain,Plain,Plain,Plain,Plain,qk),a51.stage(w,c,Attention,GateDownMLP,WideGate,Tiled,WideDownK3i,Plain,Plain,Plain,Plain),
|
|
1025
|
+
d52.stage(w,c,Delta,GateDownMLP,WideGate,Tiled,WideDownK3i,Plain,Plain,Plain,Plain,Plain,qk),d53.stage(w,c,Delta,GateDownMLP,WideGate,Tiled,WideDownK3i,Plain,Plain,Plain,Plain,Plain,qk),d54.stage(w,c,Delta,GateDownMLP,WideGate,Tiled,WideDown,Plain,Plain,Plain,Plain,Plain,qk),a55.stage(w,c,Attention,GateDownMLP,WideGate,Tiled,WideDown,Plain,Plain,Plain,Plain),
|
|
1026
|
+
d56.stage(w,c,Delta,GateDownMLP,WideGate,Tiled,WideDown,Plain,Plain,Plain,Plain,Plain,qk),d57.stage(w,c,Delta,GateDownMLP,WideGate,Tiled,WideDown,Plain,Plain,Plain,Plain,Plain,qk),d58.stage(w,c,Delta,GateDownMLP,WideGate,Tiled,WideDown,Plain,Plain,Plain,Plain,Plain,qk),a59.stage(w,c,Attention,GateDownMLP,WideGate,Tiled,WideDown,Plain,Plain,Plain,Plain),
|
|
1027
|
+
d60.stage(w,c,Delta,GateDownMLP,WideGate,Tiled,WideDown,Plain,Plain,Plain,Plain,Plain,qk),d61.stage(w,c,Delta,GateDownMLP,WideGate,Tiled,WideDown,Plain,Plain,Plain,Plain,Plain,qk),d62.stage(w,c,Delta,GateDownMLP,WideGate,Tiled,WideDownK3i,Plain,Plain,Plain,Plain,Plain,qk),a63.stage(w,c,Attention,GateDownMLP,WideGate,Tiled,WideDownK3i,Plain,Plain,Plain,Plain),
|
|
1028
|
+
];
|
|
1029
|
+
const chunkNarrow:Layers=[
|
|
1030
|
+
new Delta(w,c,`${PREFIX}.layers.0`,wide[0].mixer,mixedWideNarrow[0].mlp),new Delta(w,c,`${PREFIX}.layers.1`,wide[1].mixer,mixedWideNarrow[1].mlp),new Delta(w,c,`${PREFIX}.layers.2`,wide[2].mixer,mixedWideNarrow[2].mlp),new Attention(w,c,`${PREFIX}.layers.3`,wide[3].mixer,mixedWideNarrow[3].mlp),
|
|
1031
|
+
new Delta(w,c,`${PREFIX}.layers.4`,wide[4].mixer,mixedWideNarrow[4].mlp),new Delta(w,c,`${PREFIX}.layers.5`,wide[5].mixer,mixedWideNarrow[5].mlp),new Delta(w,c,`${PREFIX}.layers.6`,wide[6].mixer,mixedWideNarrow[6].mlp),new Attention(w,c,`${PREFIX}.layers.7`,wide[7].mixer,mixedWideNarrow[7].mlp),
|
|
1032
|
+
new Delta(w,c,`${PREFIX}.layers.8`,wide[8].mixer,mixedWideNarrow[8].mlp),new Delta(w,c,`${PREFIX}.layers.9`,wide[9].mixer,mixedWideNarrow[9].mlp),new Delta(w,c,`${PREFIX}.layers.10`,wide[10].mixer,mixedWideNarrow[10].mlp),new Attention(w,c,`${PREFIX}.layers.11`,wide[11].mixer,mixedWideNarrow[11].mlp),
|
|
1033
|
+
new Delta(w,c,`${PREFIX}.layers.12`,wide[12].mixer,mixedWideNarrow[12].mlp),new Delta(w,c,`${PREFIX}.layers.13`,wide[13].mixer,mixedWideNarrow[13].mlp),new Delta(w,c,`${PREFIX}.layers.14`,wide[14].mixer,mixedWideNarrow[14].mlp),new Attention(w,c,`${PREFIX}.layers.15`,wide[15].mixer,mixedWideNarrow[15].mlp),
|
|
1034
|
+
new Delta(w,c,`${PREFIX}.layers.16`,wide[16].mixer,mixedWideNarrow[16].mlp),new Delta(w,c,`${PREFIX}.layers.17`,wide[17].mixer,mixedWideNarrow[17].mlp),new Delta(w,c,`${PREFIX}.layers.18`,wide[18].mixer,mixedWideNarrow[18].mlp),new Attention(w,c,`${PREFIX}.layers.19`,wide[19].mixer,mixedWideNarrow[19].mlp),
|
|
1035
|
+
new Delta(w,c,`${PREFIX}.layers.20`,wide[20].mixer,mixedWideNarrow[20].mlp),new Delta(w,c,`${PREFIX}.layers.21`,wide[21].mixer,mixedWideNarrow[21].mlp),new Delta(w,c,`${PREFIX}.layers.22`,wide[22].mixer,mixedWideNarrow[22].mlp),new Attention(w,c,`${PREFIX}.layers.23`,wide[23].mixer,mixedWideNarrow[23].mlp),
|
|
1036
|
+
new Delta(w,c,`${PREFIX}.layers.24`,wide[24].mixer,mixedWideNarrow[24].mlp),new Delta(w,c,`${PREFIX}.layers.25`,wide[25].mixer,mixedWideNarrow[25].mlp),new Delta(w,c,`${PREFIX}.layers.26`,wide[26].mixer,mixedWideNarrow[26].mlp),new Attention(w,c,`${PREFIX}.layers.27`,wide[27].mixer,mixedWideNarrow[27].mlp),
|
|
1037
|
+
new Delta(w,c,`${PREFIX}.layers.28`,wide[28].mixer,mixedWideNarrow[28].mlp),new Delta(w,c,`${PREFIX}.layers.29`,wide[29].mixer,mixedWideNarrow[29].mlp),new Delta(w,c,`${PREFIX}.layers.30`,wide[30].mixer,mixedWideNarrow[30].mlp),new Attention(w,c,`${PREFIX}.layers.31`,wide[31].mixer,mixedWideNarrow[31].mlp),
|
|
1038
|
+
new Delta(w,c,`${PREFIX}.layers.32`,wide[32].mixer,mixedWideNarrow[32].mlp),new Delta(w,c,`${PREFIX}.layers.33`,wide[33].mixer,mixedWideNarrow[33].mlp),new Delta(w,c,`${PREFIX}.layers.34`,wide[34].mixer,mixedWideNarrow[34].mlp),new Attention(w,c,`${PREFIX}.layers.35`,wide[35].mixer,mixedWideNarrow[35].mlp),
|
|
1039
|
+
new Delta(w,c,`${PREFIX}.layers.36`,wide[36].mixer,mixedWideNarrow[36].mlp),new Delta(w,c,`${PREFIX}.layers.37`,wide[37].mixer,mixedWideNarrow[37].mlp),new Delta(w,c,`${PREFIX}.layers.38`,wide[38].mixer,mixedWideNarrow[38].mlp),new Attention(w,c,`${PREFIX}.layers.39`,wide[39].mixer,mixedWideNarrow[39].mlp),
|
|
1040
|
+
new Delta(w,c,`${PREFIX}.layers.40`,wide[40].mixer,mixedWideNarrow[40].mlp),new Delta(w,c,`${PREFIX}.layers.41`,wide[41].mixer,mixedWideNarrow[41].mlp),new Delta(w,c,`${PREFIX}.layers.42`,wide[42].mixer,mixedWideNarrow[42].mlp),new Attention(w,c,`${PREFIX}.layers.43`,wide[43].mixer,mixedWideNarrow[43].mlp),
|
|
1041
|
+
new Delta(w,c,`${PREFIX}.layers.44`,wide[44].mixer,mixedWideNarrow[44].mlp),new Delta(w,c,`${PREFIX}.layers.45`,wide[45].mixer,mixedWideNarrow[45].mlp),new Delta(w,c,`${PREFIX}.layers.46`,wide[46].mixer,mixedWideNarrow[46].mlp),new Attention(w,c,`${PREFIX}.layers.47`,wide[47].mixer,mixedWideNarrow[47].mlp),
|
|
1042
|
+
new Delta(w,c,`${PREFIX}.layers.48`,wide[48].mixer,mixedWideNarrow[48].mlp),new Delta(w,c,`${PREFIX}.layers.49`,wide[49].mixer,mixedWideNarrow[49].mlp),new Delta(w,c,`${PREFIX}.layers.50`,wide[50].mixer,mixedWideNarrow[50].mlp),new Attention(w,c,`${PREFIX}.layers.51`,wide[51].mixer,mixedWideNarrow[51].mlp),
|
|
1043
|
+
new Delta(w,c,`${PREFIX}.layers.52`,wide[52].mixer,mixedWideNarrow[52].mlp),new Delta(w,c,`${PREFIX}.layers.53`,wide[53].mixer,mixedWideNarrow[53].mlp),new Delta(w,c,`${PREFIX}.layers.54`,wide[54].mixer,mixedWideNarrow[54].mlp),new Attention(w,c,`${PREFIX}.layers.55`,wide[55].mixer,mixedWideNarrow[55].mlp),
|
|
1044
|
+
new Delta(w,c,`${PREFIX}.layers.56`,wide[56].mixer,mixedWideNarrow[56].mlp),new Delta(w,c,`${PREFIX}.layers.57`,wide[57].mixer,mixedWideNarrow[57].mlp),new Delta(w,c,`${PREFIX}.layers.58`,wide[58].mixer,mixedWideNarrow[58].mlp),new Attention(w,c,`${PREFIX}.layers.59`,wide[59].mixer,mixedWideNarrow[59].mlp),
|
|
1045
|
+
new Delta(w,c,`${PREFIX}.layers.60`,wide[60].mixer,mixedWideNarrow[60].mlp),new Delta(w,c,`${PREFIX}.layers.61`,wide[61].mixer,mixedWideNarrow[61].mlp),new Delta(w,c,`${PREFIX}.layers.62`,wide[62].mixer,mixedWideNarrow[62].mlp),new Attention(w,c,`${PREFIX}.layers.63`,wide[63].mixer,mixedWideNarrow[63].mlp),
|
|
1046
|
+
];
|
|
1047
|
+
const chunkFive:Layers=[
|
|
1048
|
+
new Delta(w,c,`${PREFIX}.layers.0`,wide[0].mixer,mixedWideFive[0].mlp),new Delta(w,c,`${PREFIX}.layers.1`,wide[1].mixer,mixedWideFive[1].mlp),new Delta(w,c,`${PREFIX}.layers.2`,wide[2].mixer,mixedWideFive[2].mlp),new Attention(w,c,`${PREFIX}.layers.3`,wide[3].mixer,mixedWideFive[3].mlp),
|
|
1049
|
+
new Delta(w,c,`${PREFIX}.layers.4`,wide[4].mixer,mixedWideFive[4].mlp),new Delta(w,c,`${PREFIX}.layers.5`,wide[5].mixer,mixedWideFive[5].mlp),new Delta(w,c,`${PREFIX}.layers.6`,wide[6].mixer,mixedWideFive[6].mlp),new Attention(w,c,`${PREFIX}.layers.7`,wide[7].mixer,mixedWideFive[7].mlp),
|
|
1050
|
+
new Delta(w,c,`${PREFIX}.layers.8`,wide[8].mixer,mixedWideFive[8].mlp),new Delta(w,c,`${PREFIX}.layers.9`,wide[9].mixer,mixedWideFive[9].mlp),new Delta(w,c,`${PREFIX}.layers.10`,wide[10].mixer,mixedWideFive[10].mlp),new Attention(w,c,`${PREFIX}.layers.11`,wide[11].mixer,mixedWideFive[11].mlp),
|
|
1051
|
+
new Delta(w,c,`${PREFIX}.layers.12`,wide[12].mixer,mixedWideFive[12].mlp),new Delta(w,c,`${PREFIX}.layers.13`,wide[13].mixer,mixedWideFive[13].mlp),new Delta(w,c,`${PREFIX}.layers.14`,wide[14].mixer,mixedWideFive[14].mlp),new Attention(w,c,`${PREFIX}.layers.15`,wide[15].mixer,mixedWideFive[15].mlp),
|
|
1052
|
+
new Delta(w,c,`${PREFIX}.layers.16`,wide[16].mixer,mixedWideFive[16].mlp),new Delta(w,c,`${PREFIX}.layers.17`,wide[17].mixer,mixedWideFive[17].mlp),new Delta(w,c,`${PREFIX}.layers.18`,wide[18].mixer,mixedWideFive[18].mlp),new Attention(w,c,`${PREFIX}.layers.19`,wide[19].mixer,mixedWideFive[19].mlp),
|
|
1053
|
+
new Delta(w,c,`${PREFIX}.layers.20`,wide[20].mixer,mixedWideFive[20].mlp),new Delta(w,c,`${PREFIX}.layers.21`,wide[21].mixer,mixedWideFive[21].mlp),new Delta(w,c,`${PREFIX}.layers.22`,wide[22].mixer,mixedWideFive[22].mlp),new Attention(w,c,`${PREFIX}.layers.23`,wide[23].mixer,mixedWideFive[23].mlp),
|
|
1054
|
+
new Delta(w,c,`${PREFIX}.layers.24`,wide[24].mixer,mixedWideFive[24].mlp),new Delta(w,c,`${PREFIX}.layers.25`,wide[25].mixer,mixedWideFive[25].mlp),new Delta(w,c,`${PREFIX}.layers.26`,wide[26].mixer,mixedWideFive[26].mlp),new Attention(w,c,`${PREFIX}.layers.27`,wide[27].mixer,mixedWideFive[27].mlp),
|
|
1055
|
+
new Delta(w,c,`${PREFIX}.layers.28`,wide[28].mixer,mixedWideFive[28].mlp),new Delta(w,c,`${PREFIX}.layers.29`,wide[29].mixer,mixedWideFive[29].mlp),new Delta(w,c,`${PREFIX}.layers.30`,wide[30].mixer,mixedWideFive[30].mlp),new Attention(w,c,`${PREFIX}.layers.31`,wide[31].mixer,mixedWideFive[31].mlp),
|
|
1056
|
+
new Delta(w,c,`${PREFIX}.layers.32`,wide[32].mixer,mixedWideFive[32].mlp),new Delta(w,c,`${PREFIX}.layers.33`,wide[33].mixer,mixedWideFive[33].mlp),new Delta(w,c,`${PREFIX}.layers.34`,wide[34].mixer,mixedWideFive[34].mlp),new Attention(w,c,`${PREFIX}.layers.35`,wide[35].mixer,mixedWideFive[35].mlp),
|
|
1057
|
+
new Delta(w,c,`${PREFIX}.layers.36`,wide[36].mixer,mixedWideFive[36].mlp),new Delta(w,c,`${PREFIX}.layers.37`,wide[37].mixer,mixedWideFive[37].mlp),new Delta(w,c,`${PREFIX}.layers.38`,wide[38].mixer,mixedWideFive[38].mlp),new Attention(w,c,`${PREFIX}.layers.39`,wide[39].mixer,mixedWideFive[39].mlp),
|
|
1058
|
+
new Delta(w,c,`${PREFIX}.layers.40`,wide[40].mixer,mixedWideFive[40].mlp),new Delta(w,c,`${PREFIX}.layers.41`,wide[41].mixer,mixedWideFive[41].mlp),new Delta(w,c,`${PREFIX}.layers.42`,wide[42].mixer,mixedWideFive[42].mlp),new Attention(w,c,`${PREFIX}.layers.43`,wide[43].mixer,mixedWideFive[43].mlp),
|
|
1059
|
+
new Delta(w,c,`${PREFIX}.layers.44`,wide[44].mixer,mixedWideFive[44].mlp),new Delta(w,c,`${PREFIX}.layers.45`,wide[45].mixer,mixedWideFive[45].mlp),new Delta(w,c,`${PREFIX}.layers.46`,wide[46].mixer,mixedWideFive[46].mlp),new Attention(w,c,`${PREFIX}.layers.47`,wide[47].mixer,mixedWideFive[47].mlp),
|
|
1060
|
+
new Delta(w,c,`${PREFIX}.layers.48`,wide[48].mixer,mixedWideFive[48].mlp),new Delta(w,c,`${PREFIX}.layers.49`,wide[49].mixer,mixedWideFive[49].mlp),new Delta(w,c,`${PREFIX}.layers.50`,wide[50].mixer,mixedWideFive[50].mlp),new Attention(w,c,`${PREFIX}.layers.51`,wide[51].mixer,mixedWideFive[51].mlp),
|
|
1061
|
+
new Delta(w,c,`${PREFIX}.layers.52`,wide[52].mixer,mixedWideFive[52].mlp),new Delta(w,c,`${PREFIX}.layers.53`,wide[53].mixer,mixedWideFive[53].mlp),new Delta(w,c,`${PREFIX}.layers.54`,wide[54].mixer,mixedWideFive[54].mlp),new Attention(w,c,`${PREFIX}.layers.55`,wide[55].mixer,mixedWideFive[55].mlp),
|
|
1062
|
+
new Delta(w,c,`${PREFIX}.layers.56`,wide[56].mixer,mixedWideFive[56].mlp),new Delta(w,c,`${PREFIX}.layers.57`,wide[57].mixer,mixedWideFive[57].mlp),new Delta(w,c,`${PREFIX}.layers.58`,wide[58].mixer,mixedWideFive[58].mlp),new Attention(w,c,`${PREFIX}.layers.59`,wide[59].mixer,mixedWideFive[59].mlp),
|
|
1063
|
+
new Delta(w,c,`${PREFIX}.layers.60`,wide[60].mixer,mixedWideFive[60].mlp),new Delta(w,c,`${PREFIX}.layers.61`,wide[61].mixer,mixedWideFive[61].mlp),new Delta(w,c,`${PREFIX}.layers.62`,wide[62].mixer,mixedWideFive[62].mlp),new Attention(w,c,`${PREFIX}.layers.63`,wide[63].mixer,mixedWideFive[63].mlp),
|
|
1064
|
+
];
|
|
1065
|
+
const one=this.#one,rows=this.#rows;
|
|
1066
|
+
const bindings:LayerWeights=[d0,d1,d2,a3,d4,d5,d6,a7,d8,d9,d10,a11,d12,d13,d14,a15,d16,d17,d18,a19,d20,d21,d22,a23,d24,d25,d26,a27,d28,d29,d30,a31,d32,d33,d34,a35,d36,d37,d38,a39,d40,d41,d42,a43,d44,d45,d46,a47,d48,d49,d50,a51,d52,d53,d54,a55,d56,d57,d58,a59,d60,d61,d62,a63];
|
|
1067
|
+
const shares=new AneWeightShares();
|
|
1068
|
+
this.#owned.push({dispose:()=>shares[Symbol.dispose]()});
|
|
1069
|
+
const ane2048=this.#aneAffine(w,c,bindings,qk,shares,.382,2048),ane1792=this.#aneAffine(w,c,bindings,qk,shares,.382,1792),ane1536=this.#aneAffine(w,c,bindings,qk,shares,.382,1536);
|
|
1070
|
+
const ane1280=this.#aneAffine(w,c,bindings,qk,shares,.382,1280),ane1024=this.#aneAffine(w,c,bindings,qk,shares,.382,1024),ane896=this.#aneAffine(w,c,bindings,qk,shares,.382,896),ane768=this.#aneAffine(w,c,bindings,qk,shares,.353,768);
|
|
1071
|
+
const ane704=this.#aneDense(w,c,bindings,qk,shares,.353,704),ane640=this.#aneDense(w,c,bindings,qk,shares,.353,640),ane576=this.#aneDense(w,c,bindings,qk,shares,.353,576);
|
|
1072
|
+
const ane512=this.#aneSmall(w,c,bindings,qk,shares,.353,512),ane448=this.#aneSmall(w,c,bindings,qk,shares,.353,448),ane384=this.#aneSmall(w,c,bindings,qk,shares,.353,384),ane320=this.#aneSmall(w,c,bindings,qk,shares,.353,320),ane256=this.#aneSmall(w,c,bindings,qk,shares,.353,256);
|
|
1073
|
+
const aneRanges:readonly MlxPromptRange[]=[
|
|
1074
|
+
{from:224,to:256,unit:"ane",operation:(ids,cache,capture)=>this.#queued(ane256,ids,cache,capture)},
|
|
1075
|
+
{from:257,to:320,unit:"ane",operation:(ids,cache,capture)=>this.#queued(ane320,ids,cache,capture)},
|
|
1076
|
+
{from:321,to:384,unit:"ane",operation:(ids,cache,capture)=>this.#queued(ane384,ids,cache,capture)},
|
|
1077
|
+
{from:385,to:448,unit:"ane",operation:(ids,cache,capture)=>this.#queued(ane448,ids,cache,capture)},
|
|
1078
|
+
{from:449,to:512,unit:"ane",operation:(ids,cache,capture)=>this.#queued(ane512,ids,cache,capture)},
|
|
1079
|
+
{from:513,to:576,unit:"ane",operation:(ids,cache,capture)=>this.#queued(ane576,ids,cache,capture)},
|
|
1080
|
+
{from:577,to:640,unit:"ane",operation:(ids,cache,capture)=>this.#queued(ane640,ids,cache,capture)},
|
|
1081
|
+
{from:641,to:704,unit:"ane",operation:(ids,cache,capture)=>this.#queued(ane704,ids,cache,capture)},
|
|
1082
|
+
{from:705,to:768,unit:"ane",operation:(ids,cache,capture)=>this.#queued(ane768,ids,cache,capture)},
|
|
1083
|
+
{from:769,to:896,unit:"ane",operation:(ids,cache,capture)=>this.#queued(ane896,ids,cache,capture)},
|
|
1084
|
+
{from:897,to:1024,unit:"ane",operation:(ids,cache,capture)=>this.#queued(ane1024,ids,cache,capture)},
|
|
1085
|
+
{from:1025,to:1280,unit:"ane",operation:(ids,cache,capture)=>this.#queued(ane1280,ids,cache,capture)},
|
|
1086
|
+
{from:1281,to:1536,unit:"ane",operation:(ids,cache,capture)=>this.#queued(ane1536,ids,cache,capture)},
|
|
1087
|
+
{from:1537,to:1792,unit:"ane",operation:(ids,cache,capture)=>this.#queued(ane1792,ids,cache,capture)},
|
|
1088
|
+
{from:1793,to:2048,unit:"ane",operation:(ids,cache,capture)=>this.#queued(ane2048,ids,cache,capture)},
|
|
1089
|
+
];
|
|
1090
|
+
const D1:RowsForward=(ids,cache,capture)=>this.#submitted(one,ids,cache,"decode",capture);
|
|
1091
|
+
const D23:RowsForward=(ids,cache,capture)=>this.#submitted(rows,ids,cache,"decode",capture);
|
|
1092
|
+
const D4:RowsForward=(ids,cache,capture)=>this.#submitted(mma,ids,cache,"decode",capture);
|
|
1093
|
+
const D58:RowsForward=(ids,cache,capture)=>this.#submitted(mmaFive,ids,cache,"decode",capture);
|
|
1094
|
+
const DW:RowsForward=(ids,cache,capture)=>this.#submitted(wide,ids,cache,"decode",capture);
|
|
1095
|
+
const decodeExpanded:RowsForward=(ids,cache,capture)=>this.#lazy(expanded,ids,cache,"decode",capture);
|
|
1096
|
+
this.decode=Object.freeze({
|
|
1097
|
+
1:D1,2:D23,3:D23,4:D4,5:D58,6:D58,7:D58,8:D58,9:DW,10:DW,11:DW,12:DW,13:DW,14:DW,15:DW,16:DW,
|
|
1098
|
+
17:DW,18:DW,19:DW,20:DW,21:DW,22:DW,23:DW,24:DW,25:DW,26:DW,27:DW,28:DW,29:DW,30:DW,31:DW,32:DW,
|
|
1099
|
+
33:DW,34:DW,35:DW,36:DW,37:DW,38:DW,39:DW,40:DW,41:DW,42:DW,43:DW,44:DW,45:DW,46:DW,47:DW,48:DW,
|
|
1100
|
+
49:DW,50:DW,51:DW,52:DW,53:DW,54:DW,55:DW,56:DW,57:DW,58:DW,59:DW,60:DW,61:DW,62:DW,63:DW,64:DW,
|
|
1101
|
+
65:DW,66:DW,67:DW,68:DW,69:DW,70:DW,71:DW,72:DW,73:DW,74:DW,75:DW,76:DW,77:DW,78:DW,79:DW,80:DW,
|
|
1102
|
+
81:DW,82:DW,83:DW,84:DW,85:DW,86:DW,87:DW,88:DW,89:DW,90:DW,91:DW,92:DW,93:DW,94:DW,95:DW,96:DW,
|
|
1103
|
+
97:DW,98:DW,99:DW,100:DW,101:DW,102:DW,103:DW,104:DW,105:DW,106:DW,107:DW,108:DW,109:DW,110:DW,111:DW,112:DW,
|
|
1104
|
+
113:DW,114:DW,115:DW,116:DW,117:DW,118:DW,119:DW,120:DW,121:DW,122:DW,123:DW,124:DW,125:DW,126:DW,127:DW,128:DW,
|
|
1105
|
+
129:DW,130:DW,131:DW,132:DW,133:DW,134:DW,135:DW,136:DW,137:DW,138:DW,139:DW,140:DW,141:DW,142:DW,143:DW,144:DW,
|
|
1106
|
+
145:DW,146:DW,147:DW,148:DW,149:DW,150:DW,151:DW,152:DW,153:DW,154:DW,155:DW,156:DW,157:DW,158:DW,159:DW,160:DW,
|
|
1107
|
+
161:DW,162:DW,163:DW,164:DW,165:DW,166:DW,167:DW,168:DW,169:DW,170:DW,171:DW,172:DW,173:DW,174:DW,175:DW,176:DW,
|
|
1108
|
+
177:DW,178:DW,179:DW,180:DW,181:DW,182:DW,183:DW,184:DW,185:DW,186:DW,187:DW,188:DW,189:DW,190:DW,191:DW,192:DW,
|
|
1109
|
+
193:DW,194:DW,195:DW,196:DW,197:DW,198:DW,199:DW,200:DW,201:DW,202:DW,203:DW,204:DW,205:DW,206:DW,207:DW,208:DW,
|
|
1110
|
+
209:DW,210:DW,211:DW,212:DW,213:DW,214:DW,215:DW,216:DW,217:DW,218:DW,219:DW,220:DW,221:DW,222:DW,223:DW,224:DW,
|
|
1111
|
+
225:DW,226:DW,227:DW,228:DW,229:DW,230:DW,231:DW,232:DW,233:DW,234:DW,235:DW,236:DW,237:DW,238:DW,239:DW,240:DW,
|
|
1112
|
+
241:DW,242:DW,243:DW,244:DW,245:DW,246:DW,247:DW,248:DW,249:DW,250:DW,251:DW,252:DW,253:DW,254:DW,255:DW,256:DW,
|
|
1113
|
+
|
|
1114
|
+
default: decodeExpanded,
|
|
1115
|
+
});
|
|
1116
|
+
const V1:RowsForward=(ids,cache,capture)=>this.#submitted(one,ids,cache,"window",capture);
|
|
1117
|
+
const V23:RowsForward=(ids,cache,capture)=>this.#submitted(rows,ids,cache,"window",capture);
|
|
1118
|
+
const V4:RowsForward=(ids,cache,capture)=>this.#submitted(mma,ids,cache,"window",capture);
|
|
1119
|
+
const V58:RowsForward=(ids,cache,capture)=>this.#submitted(mmaFive,ids,cache,"window",capture);
|
|
1120
|
+
const chunkNarrowWindow:RowsForward=(ids,cache,capture)=>this.#submitted(chunkNarrow,ids,cache,"window",capture);
|
|
1121
|
+
const chunkFiveWindow:RowsForward=(ids,cache,capture)=>this.#submitted(chunkFive,ids,cache,"window",capture);
|
|
1122
|
+
const VW:RowsForward=(ids,cache,capture)=>this.#submitted(wide,ids,cache,"window",capture);
|
|
1123
|
+
const windowExpanded:RowsForward=(ids,cache,capture)=>this.#lazy(expanded,ids,cache,"window",capture);
|
|
1124
|
+
this.verify=Object.freeze({
|
|
1125
|
+
1:V1,2:V23,3:V23,4:V4,5:V58,6:V58,7:V58,8:V58,9:VW,10:VW,11:VW,12:VW,13:VW,14:VW,15:VW,16:VW,
|
|
1126
|
+
17:VW,18:VW,19:VW,20:VW,21:VW,22:VW,23:VW,24:VW,25:VW,26:VW,27:VW,28:VW,29:VW,30:VW,31:VW,32:VW,
|
|
1127
|
+
33:VW,34:VW,35:VW,36:VW,37:VW,38:VW,39:VW,40:VW,41:VW,42:VW,43:VW,44:VW,45:VW,46:VW,47:VW,48:VW,
|
|
1128
|
+
49:VW,50:VW,51:VW,52:VW,53:VW,54:VW,55:VW,56:VW,57:VW,58:VW,59:VW,60:VW,61:VW,62:VW,63:VW,64:VW,
|
|
1129
|
+
65:VW,66:VW,67:VW,68:VW,69:VW,70:VW,71:VW,72:VW,73:VW,74:VW,75:VW,76:VW,77:VW,78:VW,79:VW,80:VW,
|
|
1130
|
+
81:VW,82:VW,83:VW,84:VW,85:VW,86:VW,87:VW,88:VW,89:VW,90:VW,91:VW,92:VW,93:VW,94:VW,95:VW,96:VW,
|
|
1131
|
+
97:VW,98:VW,99:VW,100:VW,101:VW,102:VW,103:VW,104:VW,105:VW,106:VW,107:VW,108:VW,109:VW,110:VW,111:VW,112:VW,
|
|
1132
|
+
113:VW,114:VW,115:VW,116:VW,117:VW,118:VW,119:VW,120:VW,121:VW,122:VW,123:VW,124:VW,125:VW,126:VW,127:VW,128:VW,
|
|
1133
|
+
129:VW,130:VW,131:VW,132:VW,133:VW,134:VW,135:VW,136:VW,137:VW,138:VW,139:VW,140:VW,141:VW,142:VW,143:VW,144:VW,
|
|
1134
|
+
145:VW,146:VW,147:VW,148:VW,149:VW,150:VW,151:VW,152:VW,153:VW,154:VW,155:VW,156:VW,157:VW,158:VW,159:VW,160:VW,
|
|
1135
|
+
161:VW,162:VW,163:VW,164:VW,165:VW,166:VW,167:VW,168:VW,169:VW,170:VW,171:VW,172:VW,173:VW,174:VW,175:VW,176:VW,
|
|
1136
|
+
177:VW,178:VW,179:VW,180:VW,181:VW,182:VW,183:VW,184:VW,185:VW,186:VW,187:VW,188:VW,189:VW,190:VW,191:VW,192:VW,
|
|
1137
|
+
193:VW,194:VW,195:VW,196:VW,197:VW,198:VW,199:VW,200:VW,201:VW,202:VW,203:VW,204:VW,205:VW,206:VW,207:VW,208:VW,
|
|
1138
|
+
209:VW,210:VW,211:VW,212:VW,213:VW,214:VW,215:VW,216:VW,217:VW,218:VW,219:VW,220:VW,221:VW,222:VW,223:VW,224:VW,
|
|
1139
|
+
225:VW,226:VW,227:VW,228:VW,229:VW,230:VW,231:VW,232:VW,233:VW,234:VW,235:VW,236:VW,237:VW,238:VW,239:VW,240:VW,
|
|
1140
|
+
241:VW,242:VW,243:VW,244:VW,245:VW,246:VW,247:VW,248:VW,249:VW,250:VW,251:VW,252:VW,253:VW,254:VW,255:VW,256:VW,
|
|
1141
|
+
|
|
1142
|
+
default: windowExpanded,
|
|
1143
|
+
});
|
|
1144
|
+
this.appendKnown=[
|
|
1145
|
+
{from:1,to:1,operation:(ids,cache)=>this.#lazy(one,ids,cache,"window",null)},
|
|
1146
|
+
{from:2,to:3,operation:(ids,cache)=>this.#lazy(rows,ids,cache,"window",null)},
|
|
1147
|
+
{from:4,to:4,operation:(ids,cache)=>this.#lazy(mma,ids,cache,"window",null)},
|
|
1148
|
+
{from:5,to:8,operation:(ids,cache)=>this.#lazy(mmaFive,ids,cache,"window",null)},
|
|
1149
|
+
{from:9,to:256,operation:(ids,cache)=>this.#lazy(wide,ids,cache,"window",null)},
|
|
1150
|
+
{from:257,to:Infinity,operation:(ids,cache)=>this.#lazy(expanded,ids,cache,"window",null)},
|
|
1151
|
+
];
|
|
1152
|
+
const rowHead=new AffineRow(this.lmHead),rowsHead=new Affine4(this.lmHead),mmaHead=new AffineMma(this.lmHead);
|
|
1153
|
+
const H1:MlxHeadForward=hidden=>rowHead.forward(hidden),H23:MlxHeadForward=hidden=>rowsHead.forward(hidden),HW:MlxHeadForward=hidden=>mmaHead.forward(hidden);
|
|
1154
|
+
const head:MlxHeadForward=hidden=>this.#head.forward(hidden);
|
|
1155
|
+
this.logits=Object.freeze({
|
|
1156
|
+
1:H1,2:H23,3:H23,4:HW,5:HW,6:HW,7:HW,8:HW,9:HW,10:HW,11:HW,12:HW,13:HW,14:HW,15:HW,16:HW,
|
|
1157
|
+
17:HW,18:HW,19:HW,20:HW,21:HW,22:HW,23:HW,24:HW,25:HW,26:HW,27:HW,28:HW,29:HW,30:HW,31:HW,32:HW,
|
|
1158
|
+
33:HW,34:HW,35:HW,36:HW,37:HW,38:HW,39:HW,40:HW,41:HW,42:HW,43:HW,44:HW,45:HW,46:HW,47:HW,48:HW,
|
|
1159
|
+
49:HW,50:HW,51:HW,52:HW,53:HW,54:HW,55:HW,56:HW,57:HW,58:HW,59:HW,60:HW,61:HW,62:HW,63:HW,64:HW,
|
|
1160
|
+
65:HW,66:HW,67:HW,68:HW,69:HW,70:HW,71:HW,72:HW,73:HW,74:HW,75:HW,76:HW,77:HW,78:HW,79:HW,80:HW,
|
|
1161
|
+
81:HW,82:HW,83:HW,84:HW,85:HW,86:HW,87:HW,88:HW,89:HW,90:HW,91:HW,92:HW,93:HW,94:HW,95:HW,96:HW,
|
|
1162
|
+
97:HW,98:HW,99:HW,100:HW,101:HW,102:HW,103:HW,104:HW,105:HW,106:HW,107:HW,108:HW,109:HW,110:HW,111:HW,112:HW,
|
|
1163
|
+
113:HW,114:HW,115:HW,116:HW,117:HW,118:HW,119:HW,120:HW,121:HW,122:HW,123:HW,124:HW,125:HW,126:HW,127:HW,128:HW,
|
|
1164
|
+
129:HW,130:HW,131:HW,132:HW,133:HW,134:HW,135:HW,136:HW,137:HW,138:HW,139:HW,140:HW,141:HW,142:HW,143:HW,144:HW,
|
|
1165
|
+
145:HW,146:HW,147:HW,148:HW,149:HW,150:HW,151:HW,152:HW,153:HW,154:HW,155:HW,156:HW,157:HW,158:HW,159:HW,160:HW,
|
|
1166
|
+
161:HW,162:HW,163:HW,164:HW,165:HW,166:HW,167:HW,168:HW,169:HW,170:HW,171:HW,172:HW,173:HW,174:HW,175:HW,176:HW,
|
|
1167
|
+
177:HW,178:HW,179:HW,180:HW,181:HW,182:HW,183:HW,184:HW,185:HW,186:HW,187:HW,188:HW,189:HW,190:HW,191:HW,192:HW,
|
|
1168
|
+
193:HW,194:HW,195:HW,196:HW,197:HW,198:HW,199:HW,200:HW,201:HW,202:HW,203:HW,204:HW,205:HW,206:HW,207:HW,208:HW,
|
|
1169
|
+
209:HW,210:HW,211:HW,212:HW,213:HW,214:HW,215:HW,216:HW,217:HW,218:HW,219:HW,220:HW,221:HW,222:HW,223:HW,224:HW,
|
|
1170
|
+
225:HW,226:HW,227:HW,228:HW,229:HW,230:HW,231:HW,232:HW,233:HW,234:HW,235:HW,236:HW,237:HW,238:HW,239:HW,240:HW,
|
|
1171
|
+
241:HW,242:HW,243:HW,244:HW,245:HW,246:HW,247:HW,248:HW,249:HW,250:HW,251:HW,252:HW,253:HW,254:HW,255:HW,256:HW,
|
|
1172
|
+
|
|
1173
|
+
default: head,
|
|
1174
|
+
});
|
|
1175
|
+
const promptSmall:RowsForward=(ids,cache,capture)=>this.#bounded(prompt512,ids,cache,"window",capture);
|
|
1176
|
+
const promptLarge:RowsForward=(ids,cache,capture)=>this.#bounded(promptDense,ids,cache,"window",capture);
|
|
1177
|
+
const gpuTail:readonly MlxPromptRange[]=[{from:1,to:1,operation:V1},{from:2,to:3,operation:V23},{from:4,to:4,operation:V4},{from:5,to:8,operation:V58},{from:9,to:256,operation:VW},{from:257,to:512,operation:promptSmall},{from:513,to:Infinity,operation:promptLarge}];
|
|
1178
|
+
this.multiSequencePrompt=gpuTail;
|
|
1179
|
+
this.multiSequencePromptChunk=[{from:1,to:4,operation:chunkNarrowWindow},{from:5,to:8,operation:chunkFiveWindow},{from:9,to:256,operation:VW},{from:257,to:512,operation:promptSmall},{from:513,to:Infinity,operation:promptLarge}];
|
|
1180
|
+
this.prompt=[{from:1,to:1,operation:V1,unit:"gpu"},{from:2,to:3,operation:V23,unit:"gpu"},{from:4,to:4,operation:V4,unit:"gpu"},{from:5,to:8,operation:V58,unit:"gpu"},{from:9,to:223,operation:VW,unit:"gpu"},...aneRanges,{from:2049,to:Infinity,operation:promptLarge,unit:"gpu"}];
|
|
1181
|
+
this.promptChunk=[{from:1,to:4,operation:chunkNarrowWindow,unit:"gpu"},{from:5,to:8,operation:chunkFiveWindow,unit:"gpu"},{from:9,to:223,operation:VW,unit:"gpu"},...aneRanges,{from:2049,to:Infinity,operation:promptLarge,unit:"gpu"}];
|
|
1182
|
+
this.sequence=[{from:1,to:1,operation:(ids,cache)=>V1(ids,cache,null)},{from:2,to:3,operation:(ids,cache)=>V23(ids,cache,null)},
|
|
1183
|
+
{from:4,to:4,operation:(ids,cache)=>V4(ids,cache,null)},{from:5,to:8,operation:(ids,cache)=>V58(ids,cache,null)},{from:9,to:256,operation:(ids,cache)=>VW(ids,cache,null)},
|
|
1184
|
+
{from:257,to:512,operation:(ids,cache)=>promptSmall(ids,cache,null)},{from:513,to:Infinity,operation:(ids,cache)=>promptLarge(ids,cache,null)}];
|
|
1185
|
+
this.mediaPrefill=[
|
|
1186
|
+
{from:1,to:4,operation:(hidden,cache,positions)=>this.#positioned(mediaStock,ops.contiguous(hidden),cache,"window",positions)},
|
|
1187
|
+
{from:5,to:8,operation:(hidden,cache,positions)=>this.#positionedBounded(mediaFive,ops.contiguous(hidden),cache,positions)},
|
|
1188
|
+
{from:9,to:32,operation:(hidden,cache,positions)=>this.#positionedBounded(mediaTiled,ops.contiguous(hidden),cache,positions)},
|
|
1189
|
+
{from:33,to:Infinity,operation:(hidden,cache,positions)=>this.#positionedBounded(mediaExpanded,ops.contiguous(hidden),cache,positions)},
|
|
1190
|
+
];
|
|
1191
|
+
this.mediaPrompt=[
|
|
1192
|
+
{from:1,to:4,operation:(ids,cache,positions)=>this.#positioned(mediaStock,this.embed.encode(ids),cache,"window",positions)},
|
|
1193
|
+
{from:5,to:8,operation:(ids,cache,positions)=>this.#positionedBounded(mediaFive,this.embed.encode(ids),cache,positions)},
|
|
1194
|
+
{from:9,to:32,operation:(ids,cache,positions)=>this.#positionedBounded(mediaTiled,this.embed.encode(ids),cache,positions)},
|
|
1195
|
+
{from:33,to:Infinity,operation:(ids,cache,positions)=>this.#positionedBounded(mediaExpanded,this.embed.encode(ids),cache,positions)},
|
|
1196
|
+
];
|
|
1197
|
+
const mediaDecodeNarrow:MlxPositionedForward=(ids,cache,positions)=>this.#positioned(mediaStock,this.embed.encode(ids),cache,"decode",positions);
|
|
1198
|
+
const mediaDecodeFive:MlxPositionedForward=(ids,cache,positions)=>this.#positioned(mediaFive,this.embed.encode(ids),cache,"decode",positions);
|
|
1199
|
+
const mediaDecodeTiled:MlxPositionedForward=(ids,cache,positions)=>this.#positioned(mediaTiled,this.embed.encode(ids),cache,"decode",positions);
|
|
1200
|
+
const mediaDecodeExpanded:MlxPositionedForward=(ids,cache,positions)=>this.#positioned(mediaExpanded,this.embed.encode(ids),cache,"decode",positions);
|
|
1201
|
+
this.mediaDecode=Object.freeze({
|
|
1202
|
+
1:mediaDecodeNarrow,2:mediaDecodeNarrow,3:mediaDecodeNarrow,4:mediaDecodeNarrow,5:mediaDecodeFive,6:mediaDecodeFive,7:mediaDecodeFive,8:mediaDecodeFive,
|
|
1203
|
+
9:mediaDecodeTiled,10:mediaDecodeTiled,11:mediaDecodeTiled,12:mediaDecodeTiled,13:mediaDecodeTiled,14:mediaDecodeTiled,15:mediaDecodeTiled,16:mediaDecodeTiled,
|
|
1204
|
+
17:mediaDecodeTiled,18:mediaDecodeTiled,19:mediaDecodeTiled,20:mediaDecodeTiled,21:mediaDecodeTiled,22:mediaDecodeTiled,23:mediaDecodeTiled,24:mediaDecodeTiled,
|
|
1205
|
+
25:mediaDecodeTiled,26:mediaDecodeTiled,27:mediaDecodeTiled,28:mediaDecodeTiled,29:mediaDecodeTiled,30:mediaDecodeTiled,31:mediaDecodeTiled,32:mediaDecodeTiled,
|
|
1206
|
+
|
|
1207
|
+
default: mediaDecodeExpanded,
|
|
1208
|
+
});
|
|
1209
|
+
this.embeddingPrompt=[
|
|
1210
|
+
{from:1,to:1,operation:(hidden,cache)=>this.#embeddingSubmitted(one,hidden,cache)},
|
|
1211
|
+
{from:2,to:3,operation:(hidden,cache)=>this.#embeddingSubmitted(rows,hidden,cache)},
|
|
1212
|
+
{from:4,to:4,operation:(hidden,cache)=>this.#embeddingSubmitted(mma,hidden,cache)},
|
|
1213
|
+
{from:5,to:8,operation:(hidden,cache)=>this.#embeddingSubmitted(mmaFive,hidden,cache)},
|
|
1214
|
+
{from:9,to:32,operation:(hidden,cache)=>this.#embeddingBounded(tiled,hidden,cache)},
|
|
1215
|
+
{from:33,to:Infinity,operation:(hidden,cache)=>this.#embeddingBounded(expanded,hidden,cache)},
|
|
1216
|
+
];
|
|
1217
|
+
this.mixed=[{from:1,to:4,operation:groups=>this.#packed(stock,groups)},
|
|
1218
|
+
{from:5,to:8,operation:groups=>this.#packed(five,groups)},
|
|
1219
|
+
{from:9,to:256,operation:groups=>this.#packed(mixedWide,groups)},
|
|
1220
|
+
{from:257,to:Infinity,operation:groups=>this.#packed(expanded,groups)}];
|
|
1221
|
+
const separateWindow:readonly MlxMixedGroupRange[]=[
|
|
1222
|
+
{from:1,to:4,operation:group=>this.#lazy(stock,group.ids,group.state,"window",group.capture)},
|
|
1223
|
+
{from:5,to:8,operation:group=>this.#bounded(five,group.ids,group.state,"window",group.capture)},
|
|
1224
|
+
{from:9,to:256,operation:group=>this.#bounded(mixedWide,group.ids,group.state,"window",group.capture)},
|
|
1225
|
+
{from:257,to:Infinity,operation:group=>this.#bounded(expanded,group.ids,group.state,"window",group.capture)},
|
|
1226
|
+
];
|
|
1227
|
+
this.mixedSeparate={decode:[
|
|
1228
|
+
{from:1,to:4,operation:group=>this.#lazy(stock,group.ids,group.state,"decode",group.capture)},
|
|
1229
|
+
{from:5,to:8,operation:group=>this.#lazy(five,group.ids,group.state,"decode",group.capture)},
|
|
1230
|
+
{from:9,to:256,operation:group=>this.#lazy(mixedWide,group.ids,group.state,"decode",group.capture)},
|
|
1231
|
+
{from:257,to:Infinity,operation:group=>decodeExpanded(group.ids,group.state,group.capture)},
|
|
1232
|
+
],window:separateWindow,verify:[
|
|
1233
|
+
{from:1,to:4,operation:group=>this.#lazy(stock,group.ids,group.state,"window",group.capture)},
|
|
1234
|
+
{from:5,to:8,operation:group=>this.#lazy(five,group.ids,group.state,"window",group.capture)},
|
|
1235
|
+
{from:9,to:256,operation:group=>this.#lazy(mixedWide,group.ids,group.state,"window",group.capture)},
|
|
1236
|
+
{from:257,to:Infinity,operation:group=>windowExpanded(group.ids,group.state,group.capture)},
|
|
1237
|
+
],packed:[
|
|
1238
|
+
{from:1,to:4,totalTo:4,operation:groups=>this.#packed(stock,groups)},
|
|
1239
|
+
{from:1,to:4,totalFrom:5,totalTo:8,operation:groups=>this.#packed(mixedUnfused,groups)},
|
|
1240
|
+
{from:5,to:8,totalTo:8,operation:groups=>this.#packed(five,groups)},
|
|
1241
|
+
{from:1,to:4,totalFrom:9,operation:groups=>this.#packed(mixedWideNarrow,groups)},
|
|
1242
|
+
{from:5,to:8,totalFrom:9,operation:groups=>this.#packed(mixedWideFive,groups)},
|
|
1243
|
+
{from:9,to:256,totalFrom:9,operation:groups=>this.#packed(mixedWide,groups)},
|
|
1244
|
+
{from:257,to:Infinity,operation:groups=>this.#packed(expanded,groups)},
|
|
1245
|
+
]};
|
|
1246
|
+
console.error(`[graph] ${this.graph}`);
|
|
1247
|
+
}
|
|
1248
|
+
|
|
1249
|
+
dispose():void { disposeResources(this.#owned.splice(0)); }
|
|
1250
|
+
#aneSmall(w:Weights,c:ModelConfig,b:LayerWeights,qk:{q:MlxArray;k:MlxArray},shares:AneWeightShares,fraction:number,capacity:number):Layers {
|
|
1251
|
+
return [
|
|
1252
|
+
b[0].withProjections(w,c,Delta,new AneBoundMlp(b[0].trellis,AneSplit,fraction,capacity,shares,4),Dense,Dense,Plain,Plain,Dense,qk),
|
|
1253
|
+
b[1].withProjections(w,c,Delta,new AneBoundMlp(b[1].trellis,AneSplitK3i,fraction,capacity,shares,4),Dense,Dense,Plain,Plain,Dense,qk),
|
|
1254
|
+
b[2].withProjections(w,c,Delta,new AneBoundMlp(b[2].trellis,AneSplitK3i,fraction,capacity,shares,4),Dense,Dense,Plain,Plain,Dense,qk),
|
|
1255
|
+
b[3].withProjections(w,c,Attention,new AneBoundMlp(b[3].trellis,AneSplitK3i,fraction,capacity,shares,4),Dense,Plain,Plain,Dense),
|
|
1256
|
+
b[4].withProjections(w,c,Delta,new AneBoundMlp(b[4].trellis,AneSplitK3i,fraction,capacity,shares,4),Dense,Dense,Plain,Plain,Dense,qk),
|
|
1257
|
+
b[5].withProjections(w,c,Delta,new AneBoundMlp(b[5].trellis,AneSplitK3i,fraction,capacity,shares,4),Dense,Dense,Plain,Plain,Dense,qk),
|
|
1258
|
+
b[6].withProjections(w,c,Delta,new AneBoundMlp(b[6].trellis,AneSplitK3i,fraction,capacity,shares,4),Dense,Dense,Plain,Plain,Dense,qk),
|
|
1259
|
+
b[7].withProjections(w,c,Attention,new AneBoundMlp(b[7].trellis,AneSplitK3i,fraction,capacity,shares,4),Dense,Plain,Plain,Dense),
|
|
1260
|
+
b[8].withProjections(w,c,Delta,new AneBoundMlp(b[8].trellis,AneSplitK3i,fraction,capacity,shares,4),Dense,Dense,Plain,Plain,Dense,qk),
|
|
1261
|
+
b[9].withProjections(w,c,Delta,new AneBoundMlp(b[9].trellis,AneSplitK3i,fraction,capacity,shares,4),Dense,Dense,Plain,Plain,Dense,qk),
|
|
1262
|
+
b[10].withProjections(w,c,Delta,new AneBoundMlp(b[10].trellis,AneSplitK3i,fraction,capacity,shares,4),Dense,Dense,Plain,Plain,Dense,qk),
|
|
1263
|
+
b[11].withProjections(w,c,Attention,new AneBoundMlp(b[11].trellis,AneSplit,fraction,capacity,shares,4),Dense,Plain,Plain,Dense),
|
|
1264
|
+
b[12].withProjections(w,c,Delta,new AneBoundMlp(b[12].trellis,AneSplitK3i,fraction,capacity,shares,4),Dense,Dense,Plain,Plain,Dense,qk),
|
|
1265
|
+
b[13].withProjections(w,c,Delta,new AneBoundMlp(b[13].trellis,AneSplit,fraction,capacity,shares,4),Dense,Dense,Plain,Plain,Dense,qk),
|
|
1266
|
+
b[14].withProjections(w,c,Delta,new AneBoundMlp(b[14].trellis,AneSplitK3i,fraction,capacity,shares,4),Dense,Dense,Plain,Plain,Dense,qk),
|
|
1267
|
+
b[15].withProjections(w,c,Attention,new AneBoundMlp(b[15].trellis,AneSplit,fraction,capacity,shares,4),Dense,Plain,Plain,Dense),
|
|
1268
|
+
b[16].withProjections(w,c,Delta,new AneBoundMlp(b[16].trellis,AneSplit,fraction,capacity,shares,4),Dense,Dense,Plain,Plain,Dense,qk),
|
|
1269
|
+
b[17].withProjections(w,c,Delta,new AneBoundMlp(b[17].trellis,AneSplit,fraction,capacity,shares,4),Dense,Dense,Plain,Plain,Dense,qk),
|
|
1270
|
+
b[18].withProjections(w,c,Delta,new AneBoundMlp(b[18].trellis,AneSplit,fraction,capacity,shares,4),Dense,Dense,Plain,Plain,Dense,qk),
|
|
1271
|
+
b[19].withProjections(w,c,Attention,new AneBoundMlp(b[19].trellis,AneSplitK3i,fraction,capacity,shares,4),Dense,Plain,Plain,Dense),
|
|
1272
|
+
b[20].withProjections(w,c,Delta,new AneBoundMlp(b[20].trellis,AneSplitK3i,fraction,capacity,shares,4),Dense,Dense,Plain,Plain,Dense,qk),
|
|
1273
|
+
b[21].withProjections(w,c,Delta,new AneBoundMlp(b[21].trellis,AneSplitK3i,fraction,capacity,shares,4),Dense,Dense,Plain,Plain,Dense,qk),
|
|
1274
|
+
b[22].withProjections(w,c,Delta,new AneBoundMlp(b[22].trellis,AneSplitK3i,fraction,capacity,shares,4),Dense,Dense,Plain,Plain,Dense,qk),
|
|
1275
|
+
b[23].withProjections(w,c,Attention,new AneBoundMlp(b[23].trellis,AneSplitK3i,fraction,capacity,shares,4),Dense,Plain,Plain,Dense),
|
|
1276
|
+
b[24].withProjections(w,c,Delta,new AneBoundMlp(b[24].trellis,AneSplitK3i,fraction,capacity,shares,4),Dense,Dense,Plain,Plain,Dense,qk),
|
|
1277
|
+
b[25].withProjections(w,c,Delta,new AneBoundMlp(b[25].trellis,AneSplitK3i,fraction,capacity,shares,4),Dense,Dense,Plain,Plain,Dense,qk),
|
|
1278
|
+
b[26].withProjections(w,c,Delta,new AneBoundMlp(b[26].trellis,AneSplitK3i,fraction,capacity,shares,4),Dense,Dense,Plain,Plain,Dense,qk),
|
|
1279
|
+
b[27].withProjections(w,c,Attention,new AneBoundMlp(b[27].trellis,AneSplitK3i,fraction,capacity,shares,4),Dense,Plain,Plain,Dense),
|
|
1280
|
+
b[28].withProjections(w,c,Delta,new AneBoundMlp(b[28].trellis,AneSplitK3i,fraction,capacity,shares,4),Dense,Dense,Plain,Plain,Dense,qk),
|
|
1281
|
+
b[29].withProjections(w,c,Delta,new AneBoundMlp(b[29].trellis,AneSplit,fraction,capacity,shares,4),Dense,Dense,Plain,Plain,Dense,qk),
|
|
1282
|
+
b[30].withProjections(w,c,Delta,new AneBoundMlp(b[30].trellis,AneSplitK3i,fraction,capacity,shares,4),Dense,Dense,Plain,Plain,Dense,qk),
|
|
1283
|
+
b[31].withProjections(w,c,Attention,new AneBoundMlp(b[31].trellis,AneSplit,fraction,capacity,shares,4),Dense,Plain,Plain,Dense),
|
|
1284
|
+
b[32].withProjections(w,c,Delta,new AneBoundMlp(b[32].trellis,AneSplit,fraction,capacity,shares,4),Dense,Dense,Plain,Plain,Dense,qk),
|
|
1285
|
+
b[33].withProjections(w,c,Delta,new AneBoundMlp(b[33].trellis,AneSplit,fraction,capacity,shares,4),Dense,Dense,Plain,Plain,Dense,qk),
|
|
1286
|
+
b[34].withProjections(w,c,Delta,new AneBoundMlp(b[34].trellis,AneSplitK3i,fraction,capacity,shares,4),Dense,Dense,Plain,Plain,Dense,qk),
|
|
1287
|
+
b[35].withProjections(w,c,Attention,new AneBoundMlp(b[35].trellis,AneSplitK3i,fraction,capacity,shares,4),Dense,Plain,Plain,Dense),
|
|
1288
|
+
b[36].withProjections(w,c,Delta,new AneBoundMlp(b[36].trellis,AneSplitK3i,fraction,capacity,shares,4),Dense,Dense,Plain,Plain,Dense,qk),
|
|
1289
|
+
b[37].withProjections(w,c,Delta,new AneBoundMlp(b[37].trellis,AneSplitK3i,fraction,capacity,shares,4),Dense,Dense,Plain,Plain,Dense,qk),
|
|
1290
|
+
b[38].withProjections(w,c,Delta,new AneBoundMlp(b[38].trellis,AneSplitK3i,fraction,capacity,shares,4),Dense,Dense,Plain,Plain,Dense,qk),
|
|
1291
|
+
b[39].withProjections(w,c,Attention,new AneBoundMlp(b[39].trellis,AneSplitK3i,fraction,capacity,shares,4),Dense,Plain,Plain,Dense),
|
|
1292
|
+
b[40].withProjections(w,c,Delta,new AneBoundMlp(b[40].trellis,AneSplitK3i,fraction,capacity,shares,4),Dense,Dense,Plain,Plain,Dense,qk),
|
|
1293
|
+
b[41].withProjections(w,c,Delta,new AneBoundMlp(b[41].trellis,AneSplitK3i,fraction,capacity,shares,4),Dense,Dense,Plain,Plain,Dense,qk),
|
|
1294
|
+
b[42].withProjections(w,c,Delta,new AneBoundMlp(b[42].trellis,AneSplitK3i,fraction,capacity,shares,4),Dense,Dense,Plain,Plain,Dense,qk),
|
|
1295
|
+
b[43].withProjections(w,c,Attention,new AneBoundMlp(b[43].trellis,AneSplitK3i,fraction,capacity,shares,4),Dense,Plain,Plain,Dense),
|
|
1296
|
+
b[44].withProjections(w,c,Delta,new AneBoundMlp(b[44].trellis,AneSplitK3i,fraction,capacity,shares,4),Dense,Dense,Plain,Plain,Dense,qk),
|
|
1297
|
+
b[45].withProjections(w,c,Delta,new AneBoundMlp(b[45].trellis,AneSplit,fraction,capacity,shares,4),Dense,Dense,Plain,Plain,Dense,qk),
|
|
1298
|
+
b[46].withProjections(w,c,Delta,new AneBoundMlp(b[46].trellis,AneSplit,fraction,capacity,shares,4),Dense,Dense,Plain,Plain,Dense,qk),
|
|
1299
|
+
b[47].withProjections(w,c,Attention,new AneBoundMlp(b[47].trellis,AneSplit,fraction,capacity,shares,4),Dense,Plain,Plain,Dense),
|
|
1300
|
+
b[48].withProjections(w,c,Delta,new AneBoundMlp(b[48].trellis,AneSplit,fraction,capacity,shares,4),Dense,Dense,Plain,Plain,Dense,qk),
|
|
1301
|
+
b[49].withProjections(w,c,Delta,new AneBoundMlp(b[49].trellis,AneSplit,fraction,capacity,shares,4),Dense,Dense,Plain,Plain,Dense,qk),
|
|
1302
|
+
b[50].withProjections(w,c,Delta,new AneBoundMlp(b[50].trellis,AneSplitK3i,fraction,capacity,shares,4),Dense,Dense,Plain,Plain,Dense,qk),
|
|
1303
|
+
b[51].withProjections(w,c,Attention,new AneBoundMlp(b[51].trellis,AneSplitK3i,fraction,capacity,shares,4),Dense,Plain,Plain,Dense),
|
|
1304
|
+
b[52].withProjections(w,c,Delta,new AneBoundMlp(b[52].trellis,AneSplitK3i,fraction,capacity,shares,4),Dense,Dense,Plain,Plain,Dense,qk),
|
|
1305
|
+
b[53].withProjections(w,c,Delta,new AneBoundMlp(b[53].trellis,AneSplitK3i,fraction,capacity,shares,4),Dense,Dense,Plain,Plain,Dense,qk),
|
|
1306
|
+
b[54].withProjections(w,c,Delta,new AneBoundMlp(b[54].trellis,AneSplit,fraction,capacity,shares,4),Dense,Dense,Plain,Plain,Dense,qk),
|
|
1307
|
+
b[55].withProjections(w,c,Attention,new AneBoundMlp(b[55].trellis,AneSplit,fraction,capacity,shares,4),Dense,Plain,Plain,Dense),
|
|
1308
|
+
b[56].withProjections(w,c,Delta,new AneBoundMlp(b[56].trellis,AneSplit,fraction,capacity,shares,4),Dense,Dense,Plain,Plain,Dense,qk),
|
|
1309
|
+
b[57].withProjections(w,c,Delta,new AneBoundMlp(b[57].trellis,AneSplit,fraction,capacity,shares,4),Dense,Dense,Plain,Plain,Dense,qk),
|
|
1310
|
+
b[58].withProjections(w,c,Delta,new AneBoundMlp(b[58].trellis,AneSplit,fraction,capacity,shares,4),Dense,Dense,Plain,Plain,Dense,qk),
|
|
1311
|
+
b[59].withProjections(w,c,Attention,new AneBoundMlp(b[59].trellis,AneSplit,fraction,capacity,shares,4),Dense,Plain,Plain,Dense),
|
|
1312
|
+
b[60].withProjections(w,c,Delta,new AneBoundMlp(b[60].trellis,AneSplit,fraction,capacity,shares,4),Dense,Dense,Plain,Plain,Dense,qk),
|
|
1313
|
+
b[61].withProjections(w,c,Delta,new AneBoundMlp(b[61].trellis,AneSplit,fraction,capacity,shares,4),Dense,Dense,Plain,Plain,Dense,qk),
|
|
1314
|
+
b[62].withProjections(w,c,Delta,new AneBoundMlp(b[62].trellis,AneSplitK3i,fraction,capacity,shares,4),Dense,Dense,Plain,Plain,Dense,qk),
|
|
1315
|
+
b[63].withProjections(w,c,Attention,new AneBoundMlp(b[63].trellis,AneSplitK3i,fraction,capacity,shares,4),Dense,Plain,Plain,Dense),
|
|
1316
|
+
];
|
|
1317
|
+
}
|
|
1318
|
+
#aneDense(w:Weights,c:ModelConfig,b:LayerWeights,qk:{q:MlxArray;k:MlxArray},shares:AneWeightShares,fraction:number,capacity:number):Layers {
|
|
1319
|
+
const small=this.#aneSmall(w,c,b,qk,shares,fraction,capacity);
|
|
1320
|
+
return [
|
|
1321
|
+
small[0],small[1],small[2],b[3].withProjections(w,c,Attention,small[3].mlp,Dense,Dense,Dense,Dense),
|
|
1322
|
+
small[4],small[5],small[6],b[7].withProjections(w,c,Attention,small[7].mlp,Dense,Dense,Dense,Dense),
|
|
1323
|
+
small[8],small[9],small[10],b[11].withProjections(w,c,Attention,small[11].mlp,Dense,Dense,Dense,Dense),
|
|
1324
|
+
small[12],small[13],small[14],b[15].withProjections(w,c,Attention,small[15].mlp,Dense,Dense,Dense,Dense),
|
|
1325
|
+
small[16],small[17],small[18],b[19].withProjections(w,c,Attention,small[19].mlp,Dense,Dense,Dense,Dense),
|
|
1326
|
+
small[20],small[21],small[22],b[23].withProjections(w,c,Attention,small[23].mlp,Dense,Dense,Dense,Dense),
|
|
1327
|
+
small[24],small[25],small[26],b[27].withProjections(w,c,Attention,small[27].mlp,Dense,Dense,Dense,Dense),
|
|
1328
|
+
small[28],small[29],small[30],b[31].withProjections(w,c,Attention,small[31].mlp,Dense,Dense,Dense,Dense),
|
|
1329
|
+
small[32],small[33],small[34],b[35].withProjections(w,c,Attention,small[35].mlp,Dense,Dense,Dense,Dense),
|
|
1330
|
+
small[36],small[37],small[38],b[39].withProjections(w,c,Attention,small[39].mlp,Dense,Dense,Dense,Dense),
|
|
1331
|
+
small[40],small[41],small[42],b[43].withProjections(w,c,Attention,small[43].mlp,Dense,Dense,Dense,Dense),
|
|
1332
|
+
small[44],small[45],small[46],b[47].withProjections(w,c,Attention,small[47].mlp,Dense,Dense,Dense,Dense),
|
|
1333
|
+
small[48],small[49],small[50],b[51].withProjections(w,c,Attention,small[51].mlp,Dense,Dense,Dense,Dense),
|
|
1334
|
+
small[52],small[53],small[54],b[55].withProjections(w,c,Attention,small[55].mlp,Dense,Dense,Dense,Dense),
|
|
1335
|
+
small[56],small[57],small[58],b[59].withProjections(w,c,Attention,small[59].mlp,Dense,Dense,Dense,Dense),
|
|
1336
|
+
small[60],small[61],small[62],b[63].withProjections(w,c,Attention,small[63].mlp,Dense,Dense,Dense,Dense),
|
|
1337
|
+
];
|
|
1338
|
+
}
|
|
1339
|
+
#aneAffine(w:Weights,c:ModelConfig,b:LayerWeights,qk:{q:MlxArray;k:MlxArray},shares:AneWeightShares,fraction:number,capacity:number):Layers {
|
|
1340
|
+
const dense=this.#aneDense(w,c,b,qk,shares,fraction,capacity);
|
|
1341
|
+
return [
|
|
1342
|
+
b[0].aneProjections(w,c,Delta,dense[0].mlp,AneAffineSplit,Dense,Plain,capacity,shares,qk),b[1].aneProjections(w,c,Delta,dense[1].mlp,AneAffineSplit,Dense,Plain,capacity,shares,qk),b[2].aneProjections(w,c,Delta,dense[2].mlp,AneAffineSplit,Dense,Plain,capacity,shares,qk),b[3].aneProjections(w,c,Attention,dense[3].mlp,AneAffineSplit,Dense,Plain,capacity,shares),
|
|
1343
|
+
b[4].aneProjections(w,c,Delta,dense[4].mlp,AneAffineSplit,Dense,Plain,capacity,shares,qk),b[5].aneProjections(w,c,Delta,dense[5].mlp,AneAffineSplit,Dense,Plain,capacity,shares,qk),b[6].aneProjections(w,c,Delta,dense[6].mlp,AneAffineSplit,Dense,Plain,capacity,shares,qk),b[7].aneProjections(w,c,Attention,dense[7].mlp,AneAffineSplit,Dense,Plain,capacity,shares),
|
|
1344
|
+
b[8].aneProjections(w,c,Delta,dense[8].mlp,AneAffineSplit,Dense,Plain,capacity,shares,qk),b[9].aneProjections(w,c,Delta,dense[9].mlp,AneAffineSplit,Dense,Plain,capacity,shares,qk),b[10].aneProjections(w,c,Delta,dense[10].mlp,AneAffineSplit,Dense,Plain,capacity,shares,qk),b[11].aneProjections(w,c,Attention,dense[11].mlp,AneAffineSplit,Dense,Plain,capacity,shares),
|
|
1345
|
+
b[12].aneProjections(w,c,Delta,dense[12].mlp,AneAffineSplit,Dense,Plain,capacity,shares,qk),b[13].aneProjections(w,c,Delta,dense[13].mlp,AneAffineSplit,Dense,Plain,capacity,shares,qk),b[14].aneProjections(w,c,Delta,dense[14].mlp,AneAffineSplit,Dense,Plain,capacity,shares,qk),b[15].aneProjections(w,c,Attention,dense[15].mlp,AneAffineSplit,Dense,Plain,capacity,shares),
|
|
1346
|
+
b[16].aneProjections(w,c,Delta,dense[16].mlp,AneAffineSplit,Dense,Plain,capacity,shares,qk),b[17].aneProjections(w,c,Delta,dense[17].mlp,AneAffineSplit,Dense,Plain,capacity,shares,qk),b[18].aneProjections(w,c,Delta,dense[18].mlp,AneAffineSplit,Dense,Plain,capacity,shares,qk),b[19].aneProjections(w,c,Attention,dense[19].mlp,AneAffineSplit,Dense,Plain,capacity,shares),
|
|
1347
|
+
b[20].aneProjections(w,c,Delta,dense[20].mlp,AneAffineSplit,Dense,Plain,capacity,shares,qk),b[21].aneProjections(w,c,Delta,dense[21].mlp,AneAffineSplit,Dense,Plain,capacity,shares,qk),b[22].aneProjections(w,c,Delta,dense[22].mlp,AneAffineSplit,Dense,Plain,capacity,shares,qk),b[23].aneProjections(w,c,Attention,dense[23].mlp,AneAffineSplit,Dense,Plain,capacity,shares),
|
|
1348
|
+
b[24].aneProjections(w,c,Delta,dense[24].mlp,AneAffineSplit,Dense,Plain,capacity,shares,qk),b[25].aneProjections(w,c,Delta,dense[25].mlp,AneAffineSplit,Dense,Plain,capacity,shares,qk),b[26].aneProjections(w,c,Delta,dense[26].mlp,AneAffineSplit,Dense,Plain,capacity,shares,qk),b[27].aneProjections(w,c,Attention,dense[27].mlp,AneAffineSplit,Dense,Plain,capacity,shares),
|
|
1349
|
+
b[28].aneProjections(w,c,Delta,dense[28].mlp,AneAffineSplit,Dense,Plain,capacity,shares,qk),b[29].aneProjections(w,c,Delta,dense[29].mlp,AneAffineSplit,Dense,Plain,capacity,shares,qk),b[30].aneProjections(w,c,Delta,dense[30].mlp,AneAffineSplit,Dense,Plain,capacity,shares,qk),b[31].aneProjections(w,c,Attention,dense[31].mlp,AneAffineSplit,Dense,Plain,capacity,shares),
|
|
1350
|
+
b[32].aneProjections(w,c,Delta,dense[32].mlp,AneAffineSplit,Dense,Plain,capacity,shares,qk),b[33].aneProjections(w,c,Delta,dense[33].mlp,AneAffineSplit,Dense,Plain,capacity,shares,qk),b[34].aneProjections(w,c,Delta,dense[34].mlp,AneAffineSplit,Dense,Plain,capacity,shares,qk),b[35].aneProjections(w,c,Attention,dense[35].mlp,AneAffineSplit,Dense,Plain,capacity,shares),
|
|
1351
|
+
b[36].aneProjections(w,c,Delta,dense[36].mlp,AneAffineSplit,Dense,Plain,capacity,shares,qk),b[37].aneProjections(w,c,Delta,dense[37].mlp,AneAffineSplit,Dense,Plain,capacity,shares,qk),b[38].aneProjections(w,c,Delta,dense[38].mlp,AneAffineSplit,Dense,Plain,capacity,shares,qk),b[39].aneProjections(w,c,Attention,dense[39].mlp,AneAffineSplit,Dense,Plain,capacity,shares),
|
|
1352
|
+
b[40].aneProjections(w,c,Delta,dense[40].mlp,AneAffineSplit,Dense,Plain,capacity,shares,qk),b[41].aneProjections(w,c,Delta,dense[41].mlp,AneAffineSplit,Dense,Plain,capacity,shares,qk),b[42].aneProjections(w,c,Delta,dense[42].mlp,AneAffineSplit,Dense,Plain,capacity,shares,qk),b[43].aneProjections(w,c,Attention,dense[43].mlp,AneAffineSplit,Dense,Plain,capacity,shares),
|
|
1353
|
+
b[44].aneProjections(w,c,Delta,dense[44].mlp,AneAffineSplit,Dense,Plain,capacity,shares,qk),b[45].aneProjections(w,c,Delta,dense[45].mlp,AneAffineSplit,Dense,Plain,capacity,shares,qk),b[46].aneProjections(w,c,Delta,dense[46].mlp,AneAffineSplit,Dense,Plain,capacity,shares,qk),b[47].aneProjections(w,c,Attention,dense[47].mlp,AneAffineSplit,Dense,Plain,capacity,shares),
|
|
1354
|
+
b[48].aneProjections(w,c,Delta,dense[48].mlp,AneAffineSplit,Dense,Plain,capacity,shares,qk),b[49].aneProjections(w,c,Delta,dense[49].mlp,AneAffineSplit,Dense,Plain,capacity,shares,qk),b[50].aneProjections(w,c,Delta,dense[50].mlp,AneAffineSplit,Dense,Plain,capacity,shares,qk),b[51].aneProjections(w,c,Attention,dense[51].mlp,AneAffineSplit,Dense,Plain,capacity,shares),
|
|
1355
|
+
b[52].aneProjections(w,c,Delta,dense[52].mlp,AneAffineSplit,Dense,Plain,capacity,shares,qk),b[53].aneProjections(w,c,Delta,dense[53].mlp,AneAffineSplit,Dense,Plain,capacity,shares,qk),b[54].aneProjections(w,c,Delta,dense[54].mlp,AneAffineSplit,Dense,Plain,capacity,shares,qk),b[55].aneProjections(w,c,Attention,dense[55].mlp,AneAffineSplit,Dense,Plain,capacity,shares),
|
|
1356
|
+
b[56].aneProjections(w,c,Delta,dense[56].mlp,AneAffineSplit,Dense,Plain,capacity,shares,qk),b[57].aneProjections(w,c,Delta,dense[57].mlp,AneAffineSplit,Dense,Plain,capacity,shares,qk),b[58].aneProjections(w,c,Delta,dense[58].mlp,AneAffineSplit,Dense,Plain,capacity,shares,qk),b[59].aneProjections(w,c,Attention,dense[59].mlp,AneAffineSplit,Dense,Plain,capacity,shares),
|
|
1357
|
+
b[60].aneProjections(w,c,Delta,dense[60].mlp,AneAffineSplit,Dense,Plain,capacity,shares,qk),b[61].aneProjections(w,c,Delta,dense[61].mlp,AneAffineSplit,Dense,Plain,capacity,shares,qk),b[62].aneProjections(w,c,Delta,dense[62].mlp,AneAffineSplit,Dense,Plain,capacity,shares,qk),b[63].aneProjections(w,c,Attention,dense[63].mlp,AneAffineSplit,Dense,Plain,capacity,shares),
|
|
1358
|
+
];
|
|
1359
|
+
}
|
|
1360
|
+
loraTargets():Map<string,QwenProjectionTarget> { return new Map(this.#targets); }
|
|
1361
|
+
bindMediaInput(input:Vision):MlxPromptInput { return bindQwenMediaInput(this,input.embeddings,input.mrope!); }
|
|
1362
|
+
mediaEncoders(modelDir:string):Promise<MediaEncoders> { return qwenMediaEncoders(this,modelDir); }
|
|
1363
|
+
draftTarget(_state:Cache[]):TargetView { return qwenDraftTarget(this,this.#residualBasis,this.#draftProjections,this.#vocabulary); }
|
|
1364
|
+
logitsFromHidden(hidden:MlxArray):MlxArray { return this.#head.forward(hidden); }
|
|
1365
|
+
#lazy(layers:Layers,ids:MlxArray,state:Cache[],phase:ReadPhase,capture:MlxLayerCapture|null):MlxArray {
|
|
1366
|
+
using scope=new GraphForwardScope(this.embed.encode(ids),capture);
|
|
1367
|
+
return this.#run(layers,state,phase,scope);
|
|
1368
|
+
}
|
|
1369
|
+
#submitted(layers:Layers,ids:MlxArray,state:Cache[],phase:ReadPhase,capture:MlxLayerCapture|null):MlxArray {
|
|
1370
|
+
using scope=new SubmittedGraphForwardScope(this.embed.encode(ids),capture,64,4);
|
|
1371
|
+
return this.#run(layers,state,phase,scope);
|
|
1372
|
+
}
|
|
1373
|
+
#queued(layers:Layers,ids:MlxArray,state:Cache[],capture:MlxLayerCapture|null):MlxArray {
|
|
1374
|
+
using scope=new QueuedGraphForwardScope(this.embed.encode(ids),capture,state);
|
|
1375
|
+
return this.#run(layers,state,"window",scope);
|
|
1376
|
+
}
|
|
1377
|
+
#embeddingSubmitted(layers:Layers,hidden:MlxArray,state:Cache[]):MlxArray {
|
|
1378
|
+
using scope=new SubmittedGraphForwardScope(ops.contiguous(hidden),null,64,4);
|
|
1379
|
+
return this.#run(layers,state,"window",scope);
|
|
1380
|
+
}
|
|
1381
|
+
#embeddingBounded(layers:Layers,hidden:MlxArray,state:Cache[]):MlxArray {
|
|
1382
|
+
using scope=new BoundedGraphForwardScope(ops.contiguous(hidden),null,state);
|
|
1383
|
+
return this.#run(layers,state,"window",scope);
|
|
1384
|
+
}
|
|
1385
|
+
#bounded(layers:Layers,ids:MlxArray,state:Cache[],phase:ReadPhase,capture:MlxLayerCapture|null):MlxArray {
|
|
1386
|
+
using scope=new BoundedGraphForwardScope(this.embed.encode(ids),capture,state);
|
|
1387
|
+
return this.#run(layers,state,phase,scope);
|
|
1388
|
+
}
|
|
1389
|
+
#positioned(layers:Layers,hidden:MlxArray,state:Cache[],phase:ReadPhase,positions:MlxArray):MlxArray {
|
|
1390
|
+
using scope=new GraphForwardScope(hidden,null);
|
|
1391
|
+
return this.#run(layers,state,phase,scope,{posIds:positions,...this.#mrope});
|
|
1392
|
+
}
|
|
1393
|
+
#positionedBounded(layers:Layers,hidden:MlxArray,state:Cache[],positions:MlxArray):MlxArray {
|
|
1394
|
+
using scope=new BoundedGraphForwardScope(hidden,null,state);
|
|
1395
|
+
return this.#run(layers,state,"window",scope,{posIds:positions,...this.#mrope});
|
|
1396
|
+
}
|
|
1397
|
+
#run(layers:Layers,state:Cache[],phase:ReadPhase,scope:GraphForwardScope,positions:MropeForwardState|null=null):MlxArray {
|
|
1398
|
+
scope.accept(0,layers[0].forward(scope.hidden,state[0]!,phase));
|
|
1399
|
+
scope.accept(1,layers[1].forward(scope.hidden,state[1]!,phase));
|
|
1400
|
+
scope.accept(2,layers[2].forward(scope.hidden,state[2]!,phase));
|
|
1401
|
+
scope.accept(3,layers[3].forward(scope.hidden,state[3]!,phase,positions));
|
|
1402
|
+
scope.accept(4,layers[4].forward(scope.hidden,state[4]!,phase));
|
|
1403
|
+
scope.accept(5,layers[5].forward(scope.hidden,state[5]!,phase));
|
|
1404
|
+
scope.accept(6,layers[6].forward(scope.hidden,state[6]!,phase));
|
|
1405
|
+
scope.accept(7,layers[7].forward(scope.hidden,state[7]!,phase,positions));
|
|
1406
|
+
scope.accept(8,layers[8].forward(scope.hidden,state[8]!,phase));
|
|
1407
|
+
scope.accept(9,layers[9].forward(scope.hidden,state[9]!,phase));
|
|
1408
|
+
scope.accept(10,layers[10].forward(scope.hidden,state[10]!,phase));
|
|
1409
|
+
scope.accept(11,layers[11].forward(scope.hidden,state[11]!,phase,positions));
|
|
1410
|
+
scope.accept(12,layers[12].forward(scope.hidden,state[12]!,phase));
|
|
1411
|
+
scope.accept(13,layers[13].forward(scope.hidden,state[13]!,phase));
|
|
1412
|
+
scope.accept(14,layers[14].forward(scope.hidden,state[14]!,phase));
|
|
1413
|
+
scope.accept(15,layers[15].forward(scope.hidden,state[15]!,phase,positions));
|
|
1414
|
+
scope.accept(16,layers[16].forward(scope.hidden,state[16]!,phase));
|
|
1415
|
+
scope.accept(17,layers[17].forward(scope.hidden,state[17]!,phase));
|
|
1416
|
+
scope.accept(18,layers[18].forward(scope.hidden,state[18]!,phase));
|
|
1417
|
+
scope.accept(19,layers[19].forward(scope.hidden,state[19]!,phase,positions));
|
|
1418
|
+
scope.accept(20,layers[20].forward(scope.hidden,state[20]!,phase));
|
|
1419
|
+
scope.accept(21,layers[21].forward(scope.hidden,state[21]!,phase));
|
|
1420
|
+
scope.accept(22,layers[22].forward(scope.hidden,state[22]!,phase));
|
|
1421
|
+
scope.accept(23,layers[23].forward(scope.hidden,state[23]!,phase,positions));
|
|
1422
|
+
scope.accept(24,layers[24].forward(scope.hidden,state[24]!,phase));
|
|
1423
|
+
scope.accept(25,layers[25].forward(scope.hidden,state[25]!,phase));
|
|
1424
|
+
scope.accept(26,layers[26].forward(scope.hidden,state[26]!,phase));
|
|
1425
|
+
scope.accept(27,layers[27].forward(scope.hidden,state[27]!,phase,positions));
|
|
1426
|
+
scope.accept(28,layers[28].forward(scope.hidden,state[28]!,phase));
|
|
1427
|
+
scope.accept(29,layers[29].forward(scope.hidden,state[29]!,phase));
|
|
1428
|
+
scope.accept(30,layers[30].forward(scope.hidden,state[30]!,phase));
|
|
1429
|
+
scope.accept(31,layers[31].forward(scope.hidden,state[31]!,phase,positions));
|
|
1430
|
+
scope.accept(32,layers[32].forward(scope.hidden,state[32]!,phase));
|
|
1431
|
+
scope.accept(33,layers[33].forward(scope.hidden,state[33]!,phase));
|
|
1432
|
+
scope.accept(34,layers[34].forward(scope.hidden,state[34]!,phase));
|
|
1433
|
+
scope.accept(35,layers[35].forward(scope.hidden,state[35]!,phase,positions));
|
|
1434
|
+
scope.accept(36,layers[36].forward(scope.hidden,state[36]!,phase));
|
|
1435
|
+
scope.accept(37,layers[37].forward(scope.hidden,state[37]!,phase));
|
|
1436
|
+
scope.accept(38,layers[38].forward(scope.hidden,state[38]!,phase));
|
|
1437
|
+
scope.accept(39,layers[39].forward(scope.hidden,state[39]!,phase,positions));
|
|
1438
|
+
scope.accept(40,layers[40].forward(scope.hidden,state[40]!,phase));
|
|
1439
|
+
scope.accept(41,layers[41].forward(scope.hidden,state[41]!,phase));
|
|
1440
|
+
scope.accept(42,layers[42].forward(scope.hidden,state[42]!,phase));
|
|
1441
|
+
scope.accept(43,layers[43].forward(scope.hidden,state[43]!,phase,positions));
|
|
1442
|
+
scope.accept(44,layers[44].forward(scope.hidden,state[44]!,phase));
|
|
1443
|
+
scope.accept(45,layers[45].forward(scope.hidden,state[45]!,phase));
|
|
1444
|
+
scope.accept(46,layers[46].forward(scope.hidden,state[46]!,phase));
|
|
1445
|
+
scope.accept(47,layers[47].forward(scope.hidden,state[47]!,phase,positions));
|
|
1446
|
+
scope.accept(48,layers[48].forward(scope.hidden,state[48]!,phase));
|
|
1447
|
+
scope.accept(49,layers[49].forward(scope.hidden,state[49]!,phase));
|
|
1448
|
+
scope.accept(50,layers[50].forward(scope.hidden,state[50]!,phase));
|
|
1449
|
+
scope.accept(51,layers[51].forward(scope.hidden,state[51]!,phase,positions));
|
|
1450
|
+
scope.accept(52,layers[52].forward(scope.hidden,state[52]!,phase));
|
|
1451
|
+
scope.accept(53,layers[53].forward(scope.hidden,state[53]!,phase));
|
|
1452
|
+
scope.accept(54,layers[54].forward(scope.hidden,state[54]!,phase));
|
|
1453
|
+
scope.accept(55,layers[55].forward(scope.hidden,state[55]!,phase,positions));
|
|
1454
|
+
scope.accept(56,layers[56].forward(scope.hidden,state[56]!,phase));
|
|
1455
|
+
scope.accept(57,layers[57].forward(scope.hidden,state[57]!,phase));
|
|
1456
|
+
scope.accept(58,layers[58].forward(scope.hidden,state[58]!,phase));
|
|
1457
|
+
scope.accept(59,layers[59].forward(scope.hidden,state[59]!,phase,positions));
|
|
1458
|
+
scope.accept(60,layers[60].forward(scope.hidden,state[60]!,phase));
|
|
1459
|
+
scope.accept(61,layers[61].forward(scope.hidden,state[61]!,phase));
|
|
1460
|
+
scope.accept(62,layers[62].forward(scope.hidden,state[62]!,phase));
|
|
1461
|
+
scope.accept(63,layers[63].forward(scope.hidden,state[63]!,phase,positions));
|
|
1462
|
+
return scope.finish(64,this.finalNorm.forward(scope.hidden));
|
|
1463
|
+
}
|
|
1464
|
+
#packed(layers:Layers,groups:readonly MlxPhaseGroup[]):MlxArray[] {
|
|
1465
|
+
using scope=new PackedMixedScope(groups,ids=>this.embed.encode(ids));
|
|
1466
|
+
scope.accept(0,(hidden,cache,phase)=>layers[0].forwardAttn(hidden,cache,phase),hidden=>layers[0].forwardMlp(hidden));
|
|
1467
|
+
scope.accept(1,(hidden,cache,phase)=>layers[1].forwardAttn(hidden,cache,phase),hidden=>layers[1].forwardMlp(hidden));
|
|
1468
|
+
scope.accept(2,(hidden,cache,phase)=>layers[2].forwardAttn(hidden,cache,phase),hidden=>layers[2].forwardMlp(hidden));
|
|
1469
|
+
scope.accept(3,(hidden,cache,phase)=>layers[3].forwardAttn(hidden,cache,phase),hidden=>layers[3].forwardMlp(hidden));
|
|
1470
|
+
scope.accept(4,(hidden,cache,phase)=>layers[4].forwardAttn(hidden,cache,phase),hidden=>layers[4].forwardMlp(hidden));
|
|
1471
|
+
scope.accept(5,(hidden,cache,phase)=>layers[5].forwardAttn(hidden,cache,phase),hidden=>layers[5].forwardMlp(hidden));
|
|
1472
|
+
scope.accept(6,(hidden,cache,phase)=>layers[6].forwardAttn(hidden,cache,phase),hidden=>layers[6].forwardMlp(hidden));
|
|
1473
|
+
scope.accept(7,(hidden,cache,phase)=>layers[7].forwardAttn(hidden,cache,phase),hidden=>layers[7].forwardMlp(hidden));
|
|
1474
|
+
scope.accept(8,(hidden,cache,phase)=>layers[8].forwardAttn(hidden,cache,phase),hidden=>layers[8].forwardMlp(hidden));
|
|
1475
|
+
scope.accept(9,(hidden,cache,phase)=>layers[9].forwardAttn(hidden,cache,phase),hidden=>layers[9].forwardMlp(hidden));
|
|
1476
|
+
scope.accept(10,(hidden,cache,phase)=>layers[10].forwardAttn(hidden,cache,phase),hidden=>layers[10].forwardMlp(hidden));
|
|
1477
|
+
scope.accept(11,(hidden,cache,phase)=>layers[11].forwardAttn(hidden,cache,phase),hidden=>layers[11].forwardMlp(hidden));
|
|
1478
|
+
scope.accept(12,(hidden,cache,phase)=>layers[12].forwardAttn(hidden,cache,phase),hidden=>layers[12].forwardMlp(hidden));
|
|
1479
|
+
scope.accept(13,(hidden,cache,phase)=>layers[13].forwardAttn(hidden,cache,phase),hidden=>layers[13].forwardMlp(hidden));
|
|
1480
|
+
scope.accept(14,(hidden,cache,phase)=>layers[14].forwardAttn(hidden,cache,phase),hidden=>layers[14].forwardMlp(hidden));
|
|
1481
|
+
scope.accept(15,(hidden,cache,phase)=>layers[15].forwardAttn(hidden,cache,phase),hidden=>layers[15].forwardMlp(hidden));
|
|
1482
|
+
scope.accept(16,(hidden,cache,phase)=>layers[16].forwardAttn(hidden,cache,phase),hidden=>layers[16].forwardMlp(hidden));
|
|
1483
|
+
scope.accept(17,(hidden,cache,phase)=>layers[17].forwardAttn(hidden,cache,phase),hidden=>layers[17].forwardMlp(hidden));
|
|
1484
|
+
scope.accept(18,(hidden,cache,phase)=>layers[18].forwardAttn(hidden,cache,phase),hidden=>layers[18].forwardMlp(hidden));
|
|
1485
|
+
scope.accept(19,(hidden,cache,phase)=>layers[19].forwardAttn(hidden,cache,phase),hidden=>layers[19].forwardMlp(hidden));
|
|
1486
|
+
scope.accept(20,(hidden,cache,phase)=>layers[20].forwardAttn(hidden,cache,phase),hidden=>layers[20].forwardMlp(hidden));
|
|
1487
|
+
scope.accept(21,(hidden,cache,phase)=>layers[21].forwardAttn(hidden,cache,phase),hidden=>layers[21].forwardMlp(hidden));
|
|
1488
|
+
scope.accept(22,(hidden,cache,phase)=>layers[22].forwardAttn(hidden,cache,phase),hidden=>layers[22].forwardMlp(hidden));
|
|
1489
|
+
scope.accept(23,(hidden,cache,phase)=>layers[23].forwardAttn(hidden,cache,phase),hidden=>layers[23].forwardMlp(hidden));
|
|
1490
|
+
scope.accept(24,(hidden,cache,phase)=>layers[24].forwardAttn(hidden,cache,phase),hidden=>layers[24].forwardMlp(hidden));
|
|
1491
|
+
scope.accept(25,(hidden,cache,phase)=>layers[25].forwardAttn(hidden,cache,phase),hidden=>layers[25].forwardMlp(hidden));
|
|
1492
|
+
scope.accept(26,(hidden,cache,phase)=>layers[26].forwardAttn(hidden,cache,phase),hidden=>layers[26].forwardMlp(hidden));
|
|
1493
|
+
scope.accept(27,(hidden,cache,phase)=>layers[27].forwardAttn(hidden,cache,phase),hidden=>layers[27].forwardMlp(hidden));
|
|
1494
|
+
scope.accept(28,(hidden,cache,phase)=>layers[28].forwardAttn(hidden,cache,phase),hidden=>layers[28].forwardMlp(hidden));
|
|
1495
|
+
scope.accept(29,(hidden,cache,phase)=>layers[29].forwardAttn(hidden,cache,phase),hidden=>layers[29].forwardMlp(hidden));
|
|
1496
|
+
scope.accept(30,(hidden,cache,phase)=>layers[30].forwardAttn(hidden,cache,phase),hidden=>layers[30].forwardMlp(hidden));
|
|
1497
|
+
scope.accept(31,(hidden,cache,phase)=>layers[31].forwardAttn(hidden,cache,phase),hidden=>layers[31].forwardMlp(hidden));
|
|
1498
|
+
scope.accept(32,(hidden,cache,phase)=>layers[32].forwardAttn(hidden,cache,phase),hidden=>layers[32].forwardMlp(hidden));
|
|
1499
|
+
scope.accept(33,(hidden,cache,phase)=>layers[33].forwardAttn(hidden,cache,phase),hidden=>layers[33].forwardMlp(hidden));
|
|
1500
|
+
scope.accept(34,(hidden,cache,phase)=>layers[34].forwardAttn(hidden,cache,phase),hidden=>layers[34].forwardMlp(hidden));
|
|
1501
|
+
scope.accept(35,(hidden,cache,phase)=>layers[35].forwardAttn(hidden,cache,phase),hidden=>layers[35].forwardMlp(hidden));
|
|
1502
|
+
scope.accept(36,(hidden,cache,phase)=>layers[36].forwardAttn(hidden,cache,phase),hidden=>layers[36].forwardMlp(hidden));
|
|
1503
|
+
scope.accept(37,(hidden,cache,phase)=>layers[37].forwardAttn(hidden,cache,phase),hidden=>layers[37].forwardMlp(hidden));
|
|
1504
|
+
scope.accept(38,(hidden,cache,phase)=>layers[38].forwardAttn(hidden,cache,phase),hidden=>layers[38].forwardMlp(hidden));
|
|
1505
|
+
scope.accept(39,(hidden,cache,phase)=>layers[39].forwardAttn(hidden,cache,phase),hidden=>layers[39].forwardMlp(hidden));
|
|
1506
|
+
scope.accept(40,(hidden,cache,phase)=>layers[40].forwardAttn(hidden,cache,phase),hidden=>layers[40].forwardMlp(hidden));
|
|
1507
|
+
scope.accept(41,(hidden,cache,phase)=>layers[41].forwardAttn(hidden,cache,phase),hidden=>layers[41].forwardMlp(hidden));
|
|
1508
|
+
scope.accept(42,(hidden,cache,phase)=>layers[42].forwardAttn(hidden,cache,phase),hidden=>layers[42].forwardMlp(hidden));
|
|
1509
|
+
scope.accept(43,(hidden,cache,phase)=>layers[43].forwardAttn(hidden,cache,phase),hidden=>layers[43].forwardMlp(hidden));
|
|
1510
|
+
scope.accept(44,(hidden,cache,phase)=>layers[44].forwardAttn(hidden,cache,phase),hidden=>layers[44].forwardMlp(hidden));
|
|
1511
|
+
scope.accept(45,(hidden,cache,phase)=>layers[45].forwardAttn(hidden,cache,phase),hidden=>layers[45].forwardMlp(hidden));
|
|
1512
|
+
scope.accept(46,(hidden,cache,phase)=>layers[46].forwardAttn(hidden,cache,phase),hidden=>layers[46].forwardMlp(hidden));
|
|
1513
|
+
scope.accept(47,(hidden,cache,phase)=>layers[47].forwardAttn(hidden,cache,phase),hidden=>layers[47].forwardMlp(hidden));
|
|
1514
|
+
scope.accept(48,(hidden,cache,phase)=>layers[48].forwardAttn(hidden,cache,phase),hidden=>layers[48].forwardMlp(hidden));
|
|
1515
|
+
scope.accept(49,(hidden,cache,phase)=>layers[49].forwardAttn(hidden,cache,phase),hidden=>layers[49].forwardMlp(hidden));
|
|
1516
|
+
scope.accept(50,(hidden,cache,phase)=>layers[50].forwardAttn(hidden,cache,phase),hidden=>layers[50].forwardMlp(hidden));
|
|
1517
|
+
scope.accept(51,(hidden,cache,phase)=>layers[51].forwardAttn(hidden,cache,phase),hidden=>layers[51].forwardMlp(hidden));
|
|
1518
|
+
scope.accept(52,(hidden,cache,phase)=>layers[52].forwardAttn(hidden,cache,phase),hidden=>layers[52].forwardMlp(hidden));
|
|
1519
|
+
scope.accept(53,(hidden,cache,phase)=>layers[53].forwardAttn(hidden,cache,phase),hidden=>layers[53].forwardMlp(hidden));
|
|
1520
|
+
scope.accept(54,(hidden,cache,phase)=>layers[54].forwardAttn(hidden,cache,phase),hidden=>layers[54].forwardMlp(hidden));
|
|
1521
|
+
scope.accept(55,(hidden,cache,phase)=>layers[55].forwardAttn(hidden,cache,phase),hidden=>layers[55].forwardMlp(hidden));
|
|
1522
|
+
scope.accept(56,(hidden,cache,phase)=>layers[56].forwardAttn(hidden,cache,phase),hidden=>layers[56].forwardMlp(hidden));
|
|
1523
|
+
scope.accept(57,(hidden,cache,phase)=>layers[57].forwardAttn(hidden,cache,phase),hidden=>layers[57].forwardMlp(hidden));
|
|
1524
|
+
scope.accept(58,(hidden,cache,phase)=>layers[58].forwardAttn(hidden,cache,phase),hidden=>layers[58].forwardMlp(hidden));
|
|
1525
|
+
scope.accept(59,(hidden,cache,phase)=>layers[59].forwardAttn(hidden,cache,phase),hidden=>layers[59].forwardMlp(hidden));
|
|
1526
|
+
scope.accept(60,(hidden,cache,phase)=>layers[60].forwardAttn(hidden,cache,phase),hidden=>layers[60].forwardMlp(hidden));
|
|
1527
|
+
scope.accept(61,(hidden,cache,phase)=>layers[61].forwardAttn(hidden,cache,phase),hidden=>layers[61].forwardMlp(hidden));
|
|
1528
|
+
scope.accept(62,(hidden,cache,phase)=>layers[62].forwardAttn(hidden,cache,phase),hidden=>layers[62].forwardMlp(hidden));
|
|
1529
|
+
scope.accept(63,(hidden,cache,phase)=>layers[63].forwardAttn(hidden,cache,phase),hidden=>layers[63].forwardMlp(hidden));
|
|
1530
|
+
return scope.finish(64,hidden=>this.finalNorm.forward(hidden));
|
|
1531
|
+
}
|
|
1532
|
+
}
|