@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 +1,1092 @@
|
|
|
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 { ResidualBasis, TargetView } from "../../contracts/mlx/draft-target";
|
|
7
|
+
import type { MlxDeclaredGraph, MlxTokenPhases, MlxLayerCapture, MlxHeadForward, MlxRowTable, MlxPromptRange, MlxCacheLayer, MlxPhaseForward, MlxSequenceRange, MlxSeparateMixed, MlxKnownRange } from "../../contracts/mlx/graph";
|
|
8
|
+
import type { MediaEncoders, MlxPromptInput, Vision, MlxPositionedForward, MlxPositionedRange, MlxPositionedEmbeddingsRange, MlxEmbeddingsRange } from "../../contracts/mlx/media";
|
|
9
|
+
import type { MropeForwardState } from "../../contracts/mlx/positions";
|
|
10
|
+
import type { GraphDeclaration } from "../../contracts/portable/graph";
|
|
11
|
+
import type { AneProgramSpec } from "../../kernels/ane/linear";
|
|
12
|
+
import { Affine3MmaLinear as Affine3Mma, Affine3RowsLinear as Affine3Rows, Affine4MmaLinear as Affine4Mma, Affine4RowsLinear as Affine4Rows } from "../../layers/affine-verify-linear";
|
|
13
|
+
import { AneAffineSplit, AneTrellisExpandedMlpSplit as AneMlp, AneTrellisExpandedMlpSplitK3i as AneMlpK3i, aneAffineSplitProgram, aneMlpSplitProgram } from "../../layers/ane-prefill-split";
|
|
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, QuantizedLinearRows as SeparateRows } from "../../layers/quantized-linear";
|
|
18
|
+
import { qwenAppendChunkSize } from "../../layers/qwen-append";
|
|
19
|
+
import { mropeInvFreq } from "../../layers/qwen-mrope";
|
|
20
|
+
import { FactoredDownK3iRow as DownK3iRow, FactoredDownK3iRows as DownK3iRows, FactoredDownRow as DownRow, FactoredDownRows as DownRows, MmaDown, MmaDownK3i } from "../../layers/trellis-down";
|
|
21
|
+
import { FactoredGateUpRow as GateRow, FactoredGateUpRows as GateRows, FactoredMixedGateUp as MixedGate, MmaGateUp as MmaGate } from "../../layers/trellis-gate-up";
|
|
22
|
+
import { TrellisFusedMLP, TrellisUnfusedMLP, TrellisReducePrompt as Reduce, TrellisScatterPrompt as Scatter, TrellisSplitKPrompt as SplitK, TrellisWidePrompt as Wide, TrellisTiledPrompt as Tiled, TrellisExpandReducePrompt as ExpandGate, TrellisExpandScatterPrompt as ExpandDown } from "../../layers/trellis-prompt";
|
|
23
|
+
import { GraphForwardScope, BoundedGraphForwardScope, SubmittedGraphForwardScope } from "../../runtime/graph-forward";
|
|
24
|
+
import { disposeResources } from "../../runtime/resources";
|
|
25
|
+
import { declareGraph } from "../capabilities";
|
|
26
|
+
import type { ReadPhase } from "../graph";
|
|
27
|
+
import { QwenDeltaLayer as Delta, QwenAttentionLayer as Attention } from "./blocks";
|
|
28
|
+
import { bindQwenMediaInput, qwenDraftTarget, qwenMediaEncoders, qwenResidualBasis } from "./media-input";
|
|
29
|
+
import { DeltaWeights, AttentionWeights, GateDownMLP, AneBoundMlp, AneWeightShares, canonicalQwenWeights, type QwenProjectionTarget } from "./written-layers";
|
|
30
|
+
|
|
31
|
+
export const QWEN38_TRELLIS_M4PRO_GRAPH = "qwen3.8-27b-trellis-3.2bpw-m4pro";
|
|
32
|
+
export const DECODE_ROWS = 8;
|
|
33
|
+
/** Exact entries through64 rows; larger rows use this file's expanded operation. */
|
|
34
|
+
export const VERIFY_ROWS = 64;
|
|
35
|
+
const PREFIX = "language_model.model";
|
|
36
|
+
const REPRESENTATIONS = { gateUpCodebook: false, mixedExactBits: false,
|
|
37
|
+
scatter: { bits: false, genericBits: false, floatCodebook: false }, widePrefill: true };
|
|
38
|
+
const ANE_FRACTION = 0.55;
|
|
39
|
+
export interface AneWidths { readonly hidden:number; readonly intermediate:number; readonly projections:readonly [readonly [number,number],readonly [number,number]] }
|
|
40
|
+
export const ANE_WIDTHS: AneWidths = {hidden:5120,intermediate:17408,projections:[[5120,6144],[6144,5120]]};
|
|
41
|
+
/** The install step fills these programs' cache; this file only loads them. */
|
|
42
|
+
export function anePrograms(widths:AneWidths=ANE_WIDTHS):AneProgramSpec[] {
|
|
43
|
+
const [[zi,zo],[oi,oo]]=widths.projections;
|
|
44
|
+
return [
|
|
45
|
+
aneMlpSplitProgram(widths.hidden,widths.intermediate,ANE_FRACTION,1024),aneAffineSplitProgram(zi,zo,ANE_FRACTION,1024),aneAffineSplitProgram(oi,oo,ANE_FRACTION,1024),
|
|
46
|
+
aneMlpSplitProgram(widths.hidden,widths.intermediate,ANE_FRACTION,1536),aneAffineSplitProgram(zi,zo,ANE_FRACTION,1536),aneAffineSplitProgram(oi,oo,ANE_FRACTION,1536),
|
|
47
|
+
aneMlpSplitProgram(widths.hidden,widths.intermediate,ANE_FRACTION,2048),aneAffineSplitProgram(zi,zo,ANE_FRACTION,2048),aneAffineSplitProgram(oi,oo,ANE_FRACTION,2048),
|
|
48
|
+
];
|
|
49
|
+
}
|
|
50
|
+
type Layers=readonly [
|
|
51
|
+
Delta,Delta,Delta,Attention,Delta,Delta,Delta,Attention,Delta,Delta,Delta,Attention,Delta,Delta,Delta,Attention,
|
|
52
|
+
Delta,Delta,Delta,Attention,Delta,Delta,Delta,Attention,Delta,Delta,Delta,Attention,Delta,Delta,Delta,Attention,
|
|
53
|
+
Delta,Delta,Delta,Attention,Delta,Delta,Delta,Attention,Delta,Delta,Delta,Attention,Delta,Delta,Delta,Attention,
|
|
54
|
+
Delta,Delta,Delta,Attention,Delta,Delta,Delta,Attention,Delta,Delta,Delta,Attention,Delta,Delta,Delta,Attention,
|
|
55
|
+
];
|
|
56
|
+
export type RowsForward=(ids:MlxArray,state:Cache[],capture:MlxLayerCapture|null)=>MlxArray;
|
|
57
|
+
|
|
58
|
+
/** A written arrangement qualified on M4 Pro. The loader recommends it;
|
|
59
|
+
* explicit selection remains the caller's choice on any machine. */
|
|
60
|
+
export class Qwen38TrellisM4Pro implements MlxDeclaredGraph,MlxTokenPhases {
|
|
61
|
+
readonly graph=QWEN38_TRELLIS_M4PRO_GRAPH;
|
|
62
|
+
readonly prefixBase=PREFIX;
|
|
63
|
+
readonly loraState=new LoraState();
|
|
64
|
+
readonly faIdx=3;
|
|
65
|
+
readonly weightsBytes:number;
|
|
66
|
+
readonly embed:QuantizedEmbedding;
|
|
67
|
+
readonly finalNorm:RMSNorm;
|
|
68
|
+
readonly lmHead:QuantizedLinear;
|
|
69
|
+
readonly graphCapabilities:GraphDeclaration=declareGraph({media:{input:"embeddings+positions",video:true},hiddenLayerTaps:true,prefill:{boundedWorkspace:true}});
|
|
70
|
+
readonly verifyRoundCosts=new Map([[1,1],[2,1.3],[3,1.46],[4,1.68],[5,1.82],[6,1.92],[7,1.93],[8,1.94]]);
|
|
71
|
+
readonly knownSpans={maxChunkSize:(state:readonly Cache[])=>qwenAppendChunkSize(state[0]!.offset)};
|
|
72
|
+
readonly cacheLayers:readonly MlxCacheLayer[]=[
|
|
73
|
+
{kind:"recurrent"},{kind:"recurrent"},{kind:"recurrent"},{kind:"full",turboQuant:"dense",kv4Head256:true},
|
|
74
|
+
{kind:"recurrent"},{kind:"recurrent"},{kind:"recurrent"},{kind:"full",turboQuant:"dense",kv4Head256:true},
|
|
75
|
+
{kind:"recurrent"},{kind:"recurrent"},{kind:"recurrent"},{kind:"full",turboQuant:"dense",kv4Head256:true},
|
|
76
|
+
{kind:"recurrent"},{kind:"recurrent"},{kind:"recurrent"},{kind:"full",turboQuant:"dense",kv4Head256:true},
|
|
77
|
+
{kind:"recurrent"},{kind:"recurrent"},{kind:"recurrent"},{kind:"full",turboQuant:"dense",kv4Head256:true},
|
|
78
|
+
{kind:"recurrent"},{kind:"recurrent"},{kind:"recurrent"},{kind:"full",turboQuant:"dense",kv4Head256:true},
|
|
79
|
+
{kind:"recurrent"},{kind:"recurrent"},{kind:"recurrent"},{kind:"full",turboQuant:"dense",kv4Head256:true},
|
|
80
|
+
{kind:"recurrent"},{kind:"recurrent"},{kind:"recurrent"},{kind:"full",turboQuant:"dense",kv4Head256:true},
|
|
81
|
+
{kind:"recurrent"},{kind:"recurrent"},{kind:"recurrent"},{kind:"full",turboQuant:"dense",kv4Head256:true},
|
|
82
|
+
{kind:"recurrent"},{kind:"recurrent"},{kind:"recurrent"},{kind:"full",turboQuant:"dense",kv4Head256:true},
|
|
83
|
+
{kind:"recurrent"},{kind:"recurrent"},{kind:"recurrent"},{kind:"full",turboQuant:"dense",kv4Head256:true},
|
|
84
|
+
{kind:"recurrent"},{kind:"recurrent"},{kind:"recurrent"},{kind:"full",turboQuant:"dense",kv4Head256:true},
|
|
85
|
+
{kind:"recurrent"},{kind:"recurrent"},{kind:"recurrent"},{kind:"full",turboQuant:"dense",kv4Head256:true},
|
|
86
|
+
{kind:"recurrent"},{kind:"recurrent"},{kind:"recurrent"},{kind:"full",turboQuant:"dense",kv4Head256:true},
|
|
87
|
+
{kind:"recurrent"},{kind:"recurrent"},{kind:"recurrent"},{kind:"full",turboQuant:"dense",kv4Head256:true},
|
|
88
|
+
{kind:"recurrent"},{kind:"recurrent"},{kind:"recurrent"},{kind:"full",turboQuant:"dense",kv4Head256:true},
|
|
89
|
+
];
|
|
90
|
+
readonly prompt:readonly MlxPromptRange[];
|
|
91
|
+
readonly promptChunk:readonly MlxPromptRange[];
|
|
92
|
+
readonly multiSequencePrompt:readonly MlxPromptRange[];
|
|
93
|
+
readonly multiSequencePromptChunk:readonly MlxPromptRange[];
|
|
94
|
+
readonly decode:MlxRowTable<RowsForward>;
|
|
95
|
+
readonly verify:MlxRowTable<RowsForward>;
|
|
96
|
+
readonly appendKnown:readonly MlxKnownRange[];
|
|
97
|
+
readonly logits:MlxRowTable<MlxHeadForward>;
|
|
98
|
+
readonly sequence:readonly MlxSequenceRange[];
|
|
99
|
+
readonly mediaPrefill:readonly MlxPositionedEmbeddingsRange[];
|
|
100
|
+
readonly mediaPrompt:readonly MlxPositionedRange[];
|
|
101
|
+
readonly mediaDecode:MlxRowTable<MlxPositionedForward>;
|
|
102
|
+
readonly embeddingPrompt:readonly MlxEmbeddingsRange[];
|
|
103
|
+
readonly mixedSeparate:MlxSeparateMixed;
|
|
104
|
+
readonly #head:Plain;
|
|
105
|
+
readonly #targets:Map<string,QwenProjectionTarget>;
|
|
106
|
+
readonly #owned:{dispose():void}[];
|
|
107
|
+
readonly #residualBasis:ResidualBasis|undefined;
|
|
108
|
+
readonly #mrope:{invFreq:MlxArray;rotaryDims:number};
|
|
109
|
+
readonly #one:Layers;
|
|
110
|
+
readonly #rows:Layers;
|
|
111
|
+
readonly #four:Layers;
|
|
112
|
+
readonly #mma:Layers;
|
|
113
|
+
readonly #wide:Layers;
|
|
114
|
+
readonly #tiled:Layers;
|
|
115
|
+
readonly #expanded:Layers;
|
|
116
|
+
readonly #ane1024:Layers;
|
|
117
|
+
readonly #ane1536:Layers;
|
|
118
|
+
readonly #ane2048:Layers;
|
|
119
|
+
readonly #stockNarrow:Layers;
|
|
120
|
+
readonly #stockFiveToEight:Layers;
|
|
121
|
+
readonly #knownNarrow:Layers;
|
|
122
|
+
readonly #knownFiveToEight:Layers;
|
|
123
|
+
readonly #knownWide:Layers;
|
|
124
|
+
readonly #knownTiled:Layers;
|
|
125
|
+
readonly #knownExpanded:Layers;
|
|
126
|
+
|
|
127
|
+
constructor(weights:Weights,readonly config:ModelConfig) {
|
|
128
|
+
const w=canonicalQwenWeights(weights),c=config,t=config.text;
|
|
129
|
+
this.weightsBytes=[...weights.shards.files.values()].reduce((bytes,file)=>bytes+file.mmap.size,0);
|
|
130
|
+
this.#residualBasis=qwenResidualBasis(this.graph,config.modelDir);
|
|
131
|
+
this.embed=QuantizedEmbedding.load(w,`${PREFIX}.embed_tokens`,c);
|
|
132
|
+
this.finalNorm=new RMSNorm(w.tensor(`${PREFIX}.norm.weight`),t.rmsNormEps);
|
|
133
|
+
this.lmHead=QuantizedLinear.load(w,"language_model.lm_head",c);
|
|
134
|
+
this.#head=new Plain(this.lmHead);
|
|
135
|
+
const invScale=Math.pow(t.linearKeyHeadDim,-0.5),qk={q:ops.filledBf16(invScale*invScale,t.linearKeyHeadDim),k:ops.filledBf16(invScale,t.linearKeyHeadDim)};
|
|
136
|
+
const rotaryDims=Math.trunc(t.headDim*t.partialRotaryFactor);
|
|
137
|
+
this.#mrope={invFreq:mropeInvFreq(rotaryDims,t.ropeParameters.full_attention!.ropeTheta),rotaryDims};
|
|
138
|
+
const shares=new AneWeightShares();
|
|
139
|
+
this.#owned=[qk.q,qk.k,this.#mrope.invFreq,{dispose:()=>shares[Symbol.dispose]()}];
|
|
140
|
+
const state=this.loraState;
|
|
141
|
+
const d0=new DeltaWeights(w,c,`${PREFIX}.layers.0`,REPRESENTATIONS,state);
|
|
142
|
+
const d1=new DeltaWeights(w,c,`${PREFIX}.layers.1`,REPRESENTATIONS,state);
|
|
143
|
+
const d2=new DeltaWeights(w,c,`${PREFIX}.layers.2`,REPRESENTATIONS,state);
|
|
144
|
+
const a3=new AttentionWeights(w,c,`${PREFIX}.layers.3`,REPRESENTATIONS,state);
|
|
145
|
+
const d4=new DeltaWeights(w,c,`${PREFIX}.layers.4`,REPRESENTATIONS,state);
|
|
146
|
+
const d5=new DeltaWeights(w,c,`${PREFIX}.layers.5`,REPRESENTATIONS,state);
|
|
147
|
+
const d6=new DeltaWeights(w,c,`${PREFIX}.layers.6`,REPRESENTATIONS,state);
|
|
148
|
+
const a7=new AttentionWeights(w,c,`${PREFIX}.layers.7`,REPRESENTATIONS,state);
|
|
149
|
+
const d8=new DeltaWeights(w,c,`${PREFIX}.layers.8`,REPRESENTATIONS,state);
|
|
150
|
+
const d9=new DeltaWeights(w,c,`${PREFIX}.layers.9`,REPRESENTATIONS,state);
|
|
151
|
+
const d10=new DeltaWeights(w,c,`${PREFIX}.layers.10`,REPRESENTATIONS,state);
|
|
152
|
+
const a11=new AttentionWeights(w,c,`${PREFIX}.layers.11`,REPRESENTATIONS,state);
|
|
153
|
+
const d12=new DeltaWeights(w,c,`${PREFIX}.layers.12`,REPRESENTATIONS,state);
|
|
154
|
+
const d13=new DeltaWeights(w,c,`${PREFIX}.layers.13`,REPRESENTATIONS,state);
|
|
155
|
+
const d14=new DeltaWeights(w,c,`${PREFIX}.layers.14`,REPRESENTATIONS,state);
|
|
156
|
+
const a15=new AttentionWeights(w,c,`${PREFIX}.layers.15`,REPRESENTATIONS,state);
|
|
157
|
+
const d16=new DeltaWeights(w,c,`${PREFIX}.layers.16`,REPRESENTATIONS,state);
|
|
158
|
+
const d17=new DeltaWeights(w,c,`${PREFIX}.layers.17`,REPRESENTATIONS,state);
|
|
159
|
+
const d18=new DeltaWeights(w,c,`${PREFIX}.layers.18`,REPRESENTATIONS,state);
|
|
160
|
+
const a19=new AttentionWeights(w,c,`${PREFIX}.layers.19`,REPRESENTATIONS,state);
|
|
161
|
+
const d20=new DeltaWeights(w,c,`${PREFIX}.layers.20`,REPRESENTATIONS,state);
|
|
162
|
+
const d21=new DeltaWeights(w,c,`${PREFIX}.layers.21`,REPRESENTATIONS,state);
|
|
163
|
+
const d22=new DeltaWeights(w,c,`${PREFIX}.layers.22`,REPRESENTATIONS,state);
|
|
164
|
+
const a23=new AttentionWeights(w,c,`${PREFIX}.layers.23`,REPRESENTATIONS,state);
|
|
165
|
+
const d24=new DeltaWeights(w,c,`${PREFIX}.layers.24`,REPRESENTATIONS,state);
|
|
166
|
+
const d25=new DeltaWeights(w,c,`${PREFIX}.layers.25`,REPRESENTATIONS,state);
|
|
167
|
+
const d26=new DeltaWeights(w,c,`${PREFIX}.layers.26`,REPRESENTATIONS,state);
|
|
168
|
+
const a27=new AttentionWeights(w,c,`${PREFIX}.layers.27`,REPRESENTATIONS,state);
|
|
169
|
+
const d28=new DeltaWeights(w,c,`${PREFIX}.layers.28`,REPRESENTATIONS,state);
|
|
170
|
+
const d29=new DeltaWeights(w,c,`${PREFIX}.layers.29`,REPRESENTATIONS,state);
|
|
171
|
+
const d30=new DeltaWeights(w,c,`${PREFIX}.layers.30`,REPRESENTATIONS,state);
|
|
172
|
+
const a31=new AttentionWeights(w,c,`${PREFIX}.layers.31`,REPRESENTATIONS,state);
|
|
173
|
+
const d32=new DeltaWeights(w,c,`${PREFIX}.layers.32`,REPRESENTATIONS,state);
|
|
174
|
+
const d33=new DeltaWeights(w,c,`${PREFIX}.layers.33`,REPRESENTATIONS,state);
|
|
175
|
+
const d34=new DeltaWeights(w,c,`${PREFIX}.layers.34`,REPRESENTATIONS,state);
|
|
176
|
+
const a35=new AttentionWeights(w,c,`${PREFIX}.layers.35`,REPRESENTATIONS,state);
|
|
177
|
+
const d36=new DeltaWeights(w,c,`${PREFIX}.layers.36`,REPRESENTATIONS,state);
|
|
178
|
+
const d37=new DeltaWeights(w,c,`${PREFIX}.layers.37`,REPRESENTATIONS,state);
|
|
179
|
+
const d38=new DeltaWeights(w,c,`${PREFIX}.layers.38`,REPRESENTATIONS,state);
|
|
180
|
+
const a39=new AttentionWeights(w,c,`${PREFIX}.layers.39`,REPRESENTATIONS,state);
|
|
181
|
+
const d40=new DeltaWeights(w,c,`${PREFIX}.layers.40`,REPRESENTATIONS,state);
|
|
182
|
+
const d41=new DeltaWeights(w,c,`${PREFIX}.layers.41`,REPRESENTATIONS,state);
|
|
183
|
+
const d42=new DeltaWeights(w,c,`${PREFIX}.layers.42`,REPRESENTATIONS,state);
|
|
184
|
+
const a43=new AttentionWeights(w,c,`${PREFIX}.layers.43`,REPRESENTATIONS,state);
|
|
185
|
+
const d44=new DeltaWeights(w,c,`${PREFIX}.layers.44`,REPRESENTATIONS,state);
|
|
186
|
+
const d45=new DeltaWeights(w,c,`${PREFIX}.layers.45`,REPRESENTATIONS,state);
|
|
187
|
+
const d46=new DeltaWeights(w,c,`${PREFIX}.layers.46`,REPRESENTATIONS,state);
|
|
188
|
+
const a47=new AttentionWeights(w,c,`${PREFIX}.layers.47`,REPRESENTATIONS,state);
|
|
189
|
+
const d48=new DeltaWeights(w,c,`${PREFIX}.layers.48`,REPRESENTATIONS,state);
|
|
190
|
+
const d49=new DeltaWeights(w,c,`${PREFIX}.layers.49`,REPRESENTATIONS,state);
|
|
191
|
+
const d50=new DeltaWeights(w,c,`${PREFIX}.layers.50`,REPRESENTATIONS,state);
|
|
192
|
+
const a51=new AttentionWeights(w,c,`${PREFIX}.layers.51`,REPRESENTATIONS,state);
|
|
193
|
+
const d52=new DeltaWeights(w,c,`${PREFIX}.layers.52`,REPRESENTATIONS,state);
|
|
194
|
+
const d53=new DeltaWeights(w,c,`${PREFIX}.layers.53`,REPRESENTATIONS,state);
|
|
195
|
+
const d54=new DeltaWeights(w,c,`${PREFIX}.layers.54`,REPRESENTATIONS,state);
|
|
196
|
+
const a55=new AttentionWeights(w,c,`${PREFIX}.layers.55`,REPRESENTATIONS,state);
|
|
197
|
+
const d56=new DeltaWeights(w,c,`${PREFIX}.layers.56`,REPRESENTATIONS,state);
|
|
198
|
+
const d57=new DeltaWeights(w,c,`${PREFIX}.layers.57`,REPRESENTATIONS,state);
|
|
199
|
+
const d58=new DeltaWeights(w,c,`${PREFIX}.layers.58`,REPRESENTATIONS,state);
|
|
200
|
+
const a59=new AttentionWeights(w,c,`${PREFIX}.layers.59`,REPRESENTATIONS,state);
|
|
201
|
+
const d60=new DeltaWeights(w,c,`${PREFIX}.layers.60`,REPRESENTATIONS,state);
|
|
202
|
+
const d61=new DeltaWeights(w,c,`${PREFIX}.layers.61`,REPRESENTATIONS,state);
|
|
203
|
+
const d62=new DeltaWeights(w,c,`${PREFIX}.layers.62`,REPRESENTATIONS,state);
|
|
204
|
+
const a63=new AttentionWeights(w,c,`${PREFIX}.layers.63`,REPRESENTATIONS,state);
|
|
205
|
+
this.#owned.push(
|
|
206
|
+
d0.trellis,d1.trellis,d2.trellis,a3.trellis,d4.trellis,d5.trellis,d6.trellis,a7.trellis,
|
|
207
|
+
d8.trellis,d9.trellis,d10.trellis,a11.trellis,d12.trellis,d13.trellis,d14.trellis,a15.trellis,
|
|
208
|
+
d16.trellis,d17.trellis,d18.trellis,a19.trellis,d20.trellis,d21.trellis,d22.trellis,a23.trellis,
|
|
209
|
+
d24.trellis,d25.trellis,d26.trellis,a27.trellis,d28.trellis,d29.trellis,d30.trellis,a31.trellis,
|
|
210
|
+
d32.trellis,d33.trellis,d34.trellis,a35.trellis,d36.trellis,d37.trellis,d38.trellis,a39.trellis,
|
|
211
|
+
d40.trellis,d41.trellis,d42.trellis,a43.trellis,d44.trellis,d45.trellis,d46.trellis,a47.trellis,
|
|
212
|
+
d48.trellis,d49.trellis,d50.trellis,a51.trellis,d52.trellis,d53.trellis,d54.trellis,a55.trellis,
|
|
213
|
+
d56.trellis,d57.trellis,d58.trellis,a59.trellis,d60.trellis,d61.trellis,d62.trellis,a63.trellis,
|
|
214
|
+
);
|
|
215
|
+
// One row: factored gate/up, with the literal mixed-width and down layout.
|
|
216
|
+
this.#one=[
|
|
217
|
+
d0.stage(w,c,Delta,GateDownMLP,GateRow,Reduce,DownRow,Plain,Plain,Plain,Plain,Plain,qk),
|
|
218
|
+
d1.stage(w,c,Delta,GateDownMLP,GateRow,Reduce,DownK3iRow,Plain,Plain,Plain,Plain,Plain,qk),
|
|
219
|
+
d2.stage(w,c,Delta,GateDownMLP,GateRow,Reduce,DownK3iRow,Plain,Plain,Plain,Plain,Plain,qk),
|
|
220
|
+
a3.stage(w,c,Attention,GateDownMLP,GateRow,Reduce,DownK3iRow,Plain,Plain,Plain,Plain),
|
|
221
|
+
d4.stage(w,c,Delta,GateDownMLP,GateRow,Reduce,DownK3iRow,Plain,Plain,Plain,Plain,Plain,qk),
|
|
222
|
+
d5.stage(w,c,Delta,GateDownMLP,GateRow,Reduce,DownK3iRow,Plain,Plain,Plain,Plain,Plain,qk),
|
|
223
|
+
d6.stage(w,c,Delta,GateDownMLP,GateRow,Reduce,DownK3iRow,Plain,Plain,Plain,Plain,Plain,qk),
|
|
224
|
+
a7.stage(w,c,Attention,GateDownMLP,GateRow,Reduce,DownK3iRow,Plain,Plain,Plain,Plain),
|
|
225
|
+
d8.stage(w,c,Delta,GateDownMLP,GateRow,Reduce,DownK3iRow,Plain,Plain,Plain,Plain,Plain,qk),
|
|
226
|
+
d9.stage(w,c,Delta,GateDownMLP,GateRow,Reduce,DownK3iRow,Plain,Plain,Plain,Plain,Plain,qk),
|
|
227
|
+
d10.stage(w,c,Delta,GateDownMLP,GateRow,Reduce,DownK3iRow,Plain,Plain,Plain,Plain,Plain,qk),
|
|
228
|
+
a11.stage(w,c,Attention,GateDownMLP,GateRow,Reduce,DownRow,Plain,Plain,Plain,Plain),
|
|
229
|
+
d12.stage(w,c,Delta,GateDownMLP,GateRow,Reduce,DownK3iRow,Plain,Plain,Plain,Plain,Plain,qk),
|
|
230
|
+
d13.stage(w,c,Delta,GateDownMLP,GateRow,Reduce,DownRow,Plain,Plain,Plain,Plain,Plain,qk),
|
|
231
|
+
d14.stage(w,c,Delta,GateDownMLP,GateRow,Reduce,DownK3iRow,Plain,Plain,Plain,Plain,Plain,qk),
|
|
232
|
+
a15.stage(w,c,Attention,GateDownMLP,MixedGate,Reduce,DownRow,Plain,Plain,Plain,Plain),
|
|
233
|
+
d16.stage(w,c,Delta,GateDownMLP,GateRow,Reduce,DownRow,Plain,Plain,Plain,Plain,Plain,qk),
|
|
234
|
+
d17.stage(w,c,Delta,GateDownMLP,GateRow,Reduce,DownRow,Plain,Plain,Plain,Plain,Plain,qk),
|
|
235
|
+
d18.stage(w,c,Delta,GateDownMLP,GateRow,Reduce,DownRow,Plain,Plain,Plain,Plain,Plain,qk),
|
|
236
|
+
a19.stage(w,c,Attention,GateDownMLP,GateRow,Reduce,DownK3iRow,Plain,Plain,Plain,Plain),
|
|
237
|
+
d20.stage(w,c,Delta,GateDownMLP,GateRow,Reduce,DownK3iRow,Plain,Plain,Plain,Plain,Plain,qk),
|
|
238
|
+
d21.stage(w,c,Delta,GateDownMLP,MixedGate,Reduce,DownK3iRow,Plain,Plain,Plain,Plain,Plain,qk),
|
|
239
|
+
d22.stage(w,c,Delta,GateDownMLP,GateRow,Reduce,DownK3iRow,Plain,Plain,Plain,Plain,Plain,qk),
|
|
240
|
+
a23.stage(w,c,Attention,GateDownMLP,GateRow,Reduce,DownK3iRow,Plain,Plain,Plain,Plain),
|
|
241
|
+
d24.stage(w,c,Delta,GateDownMLP,GateRow,Reduce,DownK3iRow,Plain,Plain,Plain,Plain,Plain,qk),
|
|
242
|
+
d25.stage(w,c,Delta,GateDownMLP,GateRow,Reduce,DownK3iRow,Plain,Plain,Plain,Plain,Plain,qk),
|
|
243
|
+
d26.stage(w,c,Delta,GateDownMLP,GateRow,Reduce,DownK3iRow,Plain,Plain,Plain,Plain,Plain,qk),
|
|
244
|
+
a27.stage(w,c,Attention,GateDownMLP,GateRow,Reduce,DownK3iRow,Plain,Plain,Plain,Plain),
|
|
245
|
+
d28.stage(w,c,Delta,GateDownMLP,GateRow,Reduce,DownK3iRow,Plain,Plain,Plain,Plain,Plain,qk),
|
|
246
|
+
d29.stage(w,c,Delta,GateDownMLP,GateRow,Reduce,DownRow,Plain,Plain,Plain,Plain,Plain,qk),
|
|
247
|
+
d30.stage(w,c,Delta,GateDownMLP,GateRow,Reduce,DownK3iRow,Plain,Plain,Plain,Plain,Plain,qk),
|
|
248
|
+
a31.stage(w,c,Attention,GateDownMLP,GateRow,Reduce,DownRow,Plain,Plain,Plain,Plain),
|
|
249
|
+
d32.stage(w,c,Delta,GateDownMLP,GateRow,Reduce,DownRow,Plain,Plain,Plain,Plain,Plain,qk),
|
|
250
|
+
d33.stage(w,c,Delta,GateDownMLP,GateRow,Reduce,DownRow,Plain,Plain,Plain,Plain,Plain,qk),
|
|
251
|
+
d34.stage(w,c,Delta,GateDownMLP,GateRow,Reduce,DownK3iRow,Plain,Plain,Plain,Plain,Plain,qk),
|
|
252
|
+
a35.stage(w,c,Attention,GateDownMLP,MixedGate,Reduce,DownK3iRow,Plain,Plain,Plain,Plain),
|
|
253
|
+
d36.stage(w,c,Delta,GateDownMLP,MixedGate,Reduce,DownK3iRow,Plain,Plain,Plain,Plain,Plain,qk),
|
|
254
|
+
d37.stage(w,c,Delta,GateDownMLP,GateRow,Reduce,DownK3iRow,Plain,Plain,Plain,Plain,Plain,qk),
|
|
255
|
+
d38.stage(w,c,Delta,GateDownMLP,GateRow,Reduce,DownK3iRow,Plain,Plain,Plain,Plain,Plain,qk),
|
|
256
|
+
a39.stage(w,c,Attention,GateDownMLP,GateRow,Reduce,DownK3iRow,Plain,Plain,Plain,Plain),
|
|
257
|
+
d40.stage(w,c,Delta,GateDownMLP,GateRow,Reduce,DownK3iRow,Plain,Plain,Plain,Plain,Plain,qk),
|
|
258
|
+
d41.stage(w,c,Delta,GateDownMLP,GateRow,Reduce,DownK3iRow,Plain,Plain,Plain,Plain,Plain,qk),
|
|
259
|
+
d42.stage(w,c,Delta,GateDownMLP,GateRow,Reduce,DownK3iRow,Plain,Plain,Plain,Plain,Plain,qk),
|
|
260
|
+
a43.stage(w,c,Attention,GateDownMLP,MixedGate,Reduce,DownK3iRow,Plain,Plain,Plain,Plain),
|
|
261
|
+
d44.stage(w,c,Delta,GateDownMLP,GateRow,Reduce,DownK3iRow,Plain,Plain,Plain,Plain,Plain,qk),
|
|
262
|
+
d45.stage(w,c,Delta,GateDownMLP,GateRow,Reduce,DownRow,Plain,Plain,Plain,Plain,Plain,qk),
|
|
263
|
+
d46.stage(w,c,Delta,GateDownMLP,GateRow,Reduce,DownRow,Plain,Plain,Plain,Plain,Plain,qk),
|
|
264
|
+
a47.stage(w,c,Attention,GateDownMLP,GateRow,Reduce,DownRow,Plain,Plain,Plain,Plain),
|
|
265
|
+
d48.stage(w,c,Delta,GateDownMLP,GateRow,Reduce,DownRow,Plain,Plain,Plain,Plain,Plain,qk),
|
|
266
|
+
d49.stage(w,c,Delta,GateDownMLP,GateRow,Reduce,DownRow,Plain,Plain,Plain,Plain,Plain,qk),
|
|
267
|
+
d50.stage(w,c,Delta,GateDownMLP,MixedGate,Reduce,DownK3iRow,Plain,Plain,Plain,Plain,Plain,qk),
|
|
268
|
+
a51.stage(w,c,Attention,GateDownMLP,MixedGate,Reduce,DownK3iRow,Plain,Plain,Plain,Plain),
|
|
269
|
+
d52.stage(w,c,Delta,GateDownMLP,GateRow,Reduce,DownK3iRow,Plain,Plain,Plain,Plain,Plain,qk),
|
|
270
|
+
d53.stage(w,c,Delta,GateDownMLP,MixedGate,Reduce,DownK3iRow,Plain,Plain,Plain,Plain,Plain,qk),
|
|
271
|
+
d54.stage(w,c,Delta,GateDownMLP,GateRow,Reduce,DownRow,Plain,Plain,Plain,Plain,Plain,qk),
|
|
272
|
+
a55.stage(w,c,Attention,GateDownMLP,GateRow,Reduce,DownRow,Plain,Plain,Plain,Plain),
|
|
273
|
+
d56.stage(w,c,Delta,GateDownMLP,MixedGate,Reduce,DownRow,Plain,Plain,Plain,Plain,Plain,qk),
|
|
274
|
+
d57.stage(w,c,Delta,GateDownMLP,GateRow,Reduce,DownRow,Plain,Plain,Plain,Plain,Plain,qk),
|
|
275
|
+
d58.stage(w,c,Delta,GateDownMLP,GateRow,Reduce,DownRow,Plain,Plain,Plain,Plain,Plain,qk),
|
|
276
|
+
a59.stage(w,c,Attention,GateDownMLP,GateRow,Reduce,DownRow,Plain,Plain,Plain,Plain),
|
|
277
|
+
d60.stage(w,c,Delta,GateDownMLP,GateRow,Reduce,DownRow,Plain,Plain,Plain,Plain,Plain,qk),
|
|
278
|
+
d61.stage(w,c,Delta,GateDownMLP,GateRow,Reduce,DownRow,Plain,Plain,Plain,Plain,Plain,qk),
|
|
279
|
+
d62.stage(w,c,Delta,GateDownMLP,MixedGate,Reduce,DownK3iRow,Plain,Plain,Plain,Plain,Plain,qk),
|
|
280
|
+
a63.stage(w,c,Attention,GateDownMLP,GateRow,Reduce,DownK3iRow,Plain,Plain,Plain,Plain),
|
|
281
|
+
];
|
|
282
|
+
// Two and three rows share factored gate/up and down, with affine row reads.
|
|
283
|
+
this.#rows=[
|
|
284
|
+
d0.stage(w,c,Delta,GateDownMLP,GateRows,Reduce,DownRows,Affine4Rows,Affine4Rows,Affine4Rows,Affine4Rows,Affine4Rows,qk),
|
|
285
|
+
d1.stage(w,c,Delta,GateDownMLP,GateRows,Reduce,DownK3iRows,Affine3Rows,Affine3Rows,Affine4Rows,Affine4Rows,Affine4Rows,qk),
|
|
286
|
+
d2.stage(w,c,Delta,GateDownMLP,GateRows,Reduce,DownK3iRows,Affine3Rows,Affine3Rows,Affine4Rows,Affine4Rows,Affine4Rows,qk),
|
|
287
|
+
a3.stage(w,c,Attention,GateDownMLP,GateRows,Reduce,DownK3iRows,Affine4Rows,Affine4Rows,Affine4Rows,Affine4Rows),
|
|
288
|
+
d4.stage(w,c,Delta,GateDownMLP,GateRows,Reduce,DownK3iRows,Affine3Rows,Affine4Rows,Affine4Rows,Affine4Rows,Affine4Rows,qk),
|
|
289
|
+
d5.stage(w,c,Delta,GateDownMLP,GateRows,Reduce,DownK3iRows,Affine3Rows,Affine4Rows,Affine4Rows,Affine4Rows,Affine4Rows,qk),
|
|
290
|
+
d6.stage(w,c,Delta,GateDownMLP,GateRows,Reduce,DownK3iRows,Affine3Rows,Affine3Rows,Affine4Rows,Affine4Rows,Affine3Rows,qk),
|
|
291
|
+
a7.stage(w,c,Attention,GateDownMLP,GateRows,Reduce,DownK3iRows,Affine4Rows,Affine4Rows,Affine4Rows,Affine4Rows),
|
|
292
|
+
d8.stage(w,c,Delta,GateDownMLP,GateRows,Reduce,DownK3iRows,Affine3Rows,Affine3Rows,Affine4Rows,Affine4Rows,Affine3Rows,qk),
|
|
293
|
+
d9.stage(w,c,Delta,GateDownMLP,GateRows,Reduce,DownK3iRows,Affine3Rows,Affine3Rows,Affine4Rows,Affine4Rows,Affine3Rows,qk),
|
|
294
|
+
d10.stage(w,c,Delta,GateDownMLP,GateRows,Reduce,DownK3iRows,Affine3Rows,Affine3Rows,Affine4Rows,Affine4Rows,Affine3Rows,qk),
|
|
295
|
+
a11.stage(w,c,Attention,GateDownMLP,GateRows,Reduce,DownRows,Affine4Rows,Affine4Rows,Affine4Rows,Affine4Rows),
|
|
296
|
+
d12.stage(w,c,Delta,GateDownMLP,GateRows,Reduce,DownK3iRows,Affine3Rows,Affine3Rows,Affine4Rows,Affine4Rows,Affine3Rows,qk),
|
|
297
|
+
d13.stage(w,c,Delta,GateDownMLP,GateRows,Reduce,DownRows,Affine3Rows,Affine3Rows,Affine4Rows,Affine4Rows,Affine3Rows,qk),
|
|
298
|
+
d14.stage(w,c,Delta,GateDownMLP,GateRows,Reduce,DownK3iRows,Affine3Rows,Affine3Rows,Affine4Rows,Affine4Rows,Affine3Rows,qk),
|
|
299
|
+
a15.stage(w,c,Attention,GateDownMLP,MixedGate,Reduce,DownRows,Affine4Rows,Affine4Rows,Affine4Rows,Affine4Rows),
|
|
300
|
+
d16.stage(w,c,Delta,GateDownMLP,GateRows,Reduce,DownRows,Affine3Rows,Affine3Rows,Affine4Rows,Affine4Rows,Affine3Rows,qk),
|
|
301
|
+
d17.stage(w,c,Delta,GateDownMLP,GateRows,Reduce,DownRows,Affine3Rows,Affine3Rows,Affine4Rows,Affine4Rows,Affine3Rows,qk),
|
|
302
|
+
d18.stage(w,c,Delta,GateDownMLP,GateRows,Reduce,DownRows,Affine3Rows,Affine3Rows,Affine4Rows,Affine4Rows,Affine3Rows,qk),
|
|
303
|
+
a19.stage(w,c,Attention,GateDownMLP,GateRows,Reduce,DownK3iRows,Affine4Rows,Affine4Rows,Affine4Rows,Affine4Rows),
|
|
304
|
+
d20.stage(w,c,Delta,GateDownMLP,GateRows,Reduce,DownK3iRows,Affine3Rows,Affine3Rows,Affine4Rows,Affine4Rows,Affine3Rows,qk),
|
|
305
|
+
d21.stage(w,c,Delta,GateDownMLP,MixedGate,Reduce,DownK3iRows,Affine3Rows,Affine3Rows,Affine4Rows,Affine4Rows,Affine3Rows,qk),
|
|
306
|
+
d22.stage(w,c,Delta,GateDownMLP,GateRows,Reduce,DownK3iRows,Affine3Rows,Affine3Rows,Affine4Rows,Affine4Rows,Affine3Rows,qk),
|
|
307
|
+
a23.stage(w,c,Attention,GateDownMLP,GateRows,Reduce,DownK3iRows,Affine4Rows,Affine4Rows,Affine4Rows,Affine4Rows),
|
|
308
|
+
d24.stage(w,c,Delta,GateDownMLP,GateRows,Reduce,DownK3iRows,Affine3Rows,Affine3Rows,Affine4Rows,Affine4Rows,Affine3Rows,qk),
|
|
309
|
+
d25.stage(w,c,Delta,GateDownMLP,GateRows,Reduce,DownK3iRows,Affine3Rows,Affine4Rows,Affine4Rows,Affine4Rows,Affine3Rows,qk),
|
|
310
|
+
d26.stage(w,c,Delta,GateDownMLP,GateRows,Reduce,DownK3iRows,Affine3Rows,Affine3Rows,Affine4Rows,Affine4Rows,Affine3Rows,qk),
|
|
311
|
+
a27.stage(w,c,Attention,GateDownMLP,GateRows,Reduce,DownK3iRows,Affine4Rows,Affine4Rows,Affine4Rows,Affine4Rows),
|
|
312
|
+
d28.stage(w,c,Delta,GateDownMLP,GateRows,Reduce,DownK3iRows,Affine3Rows,Affine3Rows,Affine4Rows,Affine4Rows,Affine3Rows,qk),
|
|
313
|
+
d29.stage(w,c,Delta,GateDownMLP,GateRows,Reduce,DownRows,Affine3Rows,Affine3Rows,Affine4Rows,Affine4Rows,Affine3Rows,qk),
|
|
314
|
+
d30.stage(w,c,Delta,GateDownMLP,GateRows,Reduce,DownK3iRows,Affine3Rows,Affine3Rows,Affine4Rows,Affine4Rows,Affine3Rows,qk),
|
|
315
|
+
a31.stage(w,c,Attention,GateDownMLP,GateRows,Reduce,DownRows,Affine4Rows,Affine4Rows,Affine4Rows,Affine4Rows),
|
|
316
|
+
d32.stage(w,c,Delta,GateDownMLP,GateRows,Reduce,DownRows,Affine3Rows,Affine3Rows,Affine4Rows,Affine4Rows,Affine3Rows,qk),
|
|
317
|
+
d33.stage(w,c,Delta,GateDownMLP,GateRows,Reduce,DownRows,Affine3Rows,Affine3Rows,Affine4Rows,Affine4Rows,Affine3Rows,qk),
|
|
318
|
+
d34.stage(w,c,Delta,GateDownMLP,GateRows,Reduce,DownK3iRows,Affine3Rows,Affine3Rows,Affine4Rows,Affine4Rows,Affine3Rows,qk),
|
|
319
|
+
a35.stage(w,c,Attention,GateDownMLP,MixedGate,Reduce,DownK3iRows,Affine4Rows,Affine4Rows,Affine4Rows,Affine4Rows),
|
|
320
|
+
d36.stage(w,c,Delta,GateDownMLP,MixedGate,Reduce,DownK3iRows,Affine3Rows,Affine3Rows,Affine4Rows,Affine4Rows,Affine3Rows,qk),
|
|
321
|
+
d37.stage(w,c,Delta,GateDownMLP,GateRows,Reduce,DownK3iRows,Affine3Rows,Affine3Rows,Affine4Rows,Affine4Rows,Affine3Rows,qk),
|
|
322
|
+
d38.stage(w,c,Delta,GateDownMLP,GateRows,Reduce,DownK3iRows,Affine3Rows,Affine3Rows,Affine4Rows,Affine4Rows,Affine3Rows,qk),
|
|
323
|
+
a39.stage(w,c,Attention,GateDownMLP,GateRows,Reduce,DownK3iRows,Affine4Rows,Affine4Rows,Affine4Rows,Affine4Rows),
|
|
324
|
+
d40.stage(w,c,Delta,GateDownMLP,GateRows,Reduce,DownK3iRows,Affine3Rows,Affine3Rows,Affine4Rows,Affine4Rows,Affine3Rows,qk),
|
|
325
|
+
d41.stage(w,c,Delta,GateDownMLP,GateRows,Reduce,DownK3iRows,Affine3Rows,Affine3Rows,Affine4Rows,Affine4Rows,Affine3Rows,qk),
|
|
326
|
+
d42.stage(w,c,Delta,GateDownMLP,GateRows,Reduce,DownK3iRows,Affine3Rows,Affine3Rows,Affine4Rows,Affine4Rows,Affine3Rows,qk),
|
|
327
|
+
a43.stage(w,c,Attention,GateDownMLP,MixedGate,Reduce,DownK3iRows,Affine4Rows,Affine4Rows,Affine4Rows,Affine4Rows),
|
|
328
|
+
d44.stage(w,c,Delta,GateDownMLP,GateRows,Reduce,DownK3iRows,Affine3Rows,Affine3Rows,Affine4Rows,Affine4Rows,Affine3Rows,qk),
|
|
329
|
+
d45.stage(w,c,Delta,GateDownMLP,GateRows,Reduce,DownRows,Affine3Rows,Affine3Rows,Affine4Rows,Affine4Rows,Affine3Rows,qk),
|
|
330
|
+
d46.stage(w,c,Delta,GateDownMLP,GateRows,Reduce,DownRows,Affine3Rows,Affine3Rows,Affine4Rows,Affine4Rows,Affine3Rows,qk),
|
|
331
|
+
a47.stage(w,c,Attention,GateDownMLP,GateRows,Reduce,DownRows,Affine4Rows,Affine4Rows,Affine4Rows,Affine4Rows),
|
|
332
|
+
d48.stage(w,c,Delta,GateDownMLP,GateRows,Reduce,DownRows,Affine3Rows,Affine3Rows,Affine4Rows,Affine4Rows,Affine3Rows,qk),
|
|
333
|
+
d49.stage(w,c,Delta,GateDownMLP,GateRows,Reduce,DownRows,Affine3Rows,Affine3Rows,Affine4Rows,Affine4Rows,Affine3Rows,qk),
|
|
334
|
+
d50.stage(w,c,Delta,GateDownMLP,MixedGate,Reduce,DownK3iRows,Affine3Rows,Affine3Rows,Affine4Rows,Affine4Rows,Affine3Rows,qk),
|
|
335
|
+
a51.stage(w,c,Attention,GateDownMLP,MixedGate,Reduce,DownK3iRows,Affine4Rows,Affine4Rows,Affine4Rows,Affine4Rows),
|
|
336
|
+
d52.stage(w,c,Delta,GateDownMLP,GateRows,Reduce,DownK3iRows,Affine3Rows,Affine3Rows,Affine4Rows,Affine4Rows,Affine3Rows,qk),
|
|
337
|
+
d53.stage(w,c,Delta,GateDownMLP,MixedGate,Reduce,DownK3iRows,Affine3Rows,Affine4Rows,Affine4Rows,Affine4Rows,Affine3Rows,qk),
|
|
338
|
+
d54.stage(w,c,Delta,GateDownMLP,GateRows,Reduce,DownRows,Affine4Rows,Affine4Rows,Affine4Rows,Affine4Rows,Affine4Rows,qk),
|
|
339
|
+
a55.stage(w,c,Attention,GateDownMLP,GateRows,Reduce,DownRows,Affine4Rows,Affine4Rows,Affine4Rows,Affine4Rows),
|
|
340
|
+
d56.stage(w,c,Delta,GateDownMLP,MixedGate,Reduce,DownRows,Affine4Rows,Affine4Rows,Affine4Rows,Affine4Rows,Affine4Rows,qk),
|
|
341
|
+
d57.stage(w,c,Delta,GateDownMLP,GateRows,Reduce,DownRows,Affine4Rows,Affine4Rows,Affine4Rows,Affine4Rows,Affine4Rows,qk),
|
|
342
|
+
d58.stage(w,c,Delta,GateDownMLP,GateRows,Reduce,DownRows,Affine4Rows,Affine3Rows,Affine4Rows,Affine4Rows,Affine3Rows,qk),
|
|
343
|
+
a59.stage(w,c,Attention,GateDownMLP,GateRows,Reduce,DownRows,Affine4Rows,Affine4Rows,Affine4Rows,Affine4Rows),
|
|
344
|
+
d60.stage(w,c,Delta,GateDownMLP,GateRows,Reduce,DownRows,Affine4Rows,Affine4Rows,Affine4Rows,Affine4Rows,Affine3Rows,qk),
|
|
345
|
+
d61.stage(w,c,Delta,GateDownMLP,GateRows,Reduce,DownRows,Affine4Rows,Affine4Rows,Affine4Rows,Affine4Rows,Affine4Rows,qk),
|
|
346
|
+
d62.stage(w,c,Delta,GateDownMLP,MixedGate,Reduce,DownK3iRows,Affine4Rows,Affine4Rows,Affine4Rows,Affine4Rows,Affine3Rows,qk),
|
|
347
|
+
a63.stage(w,c,Attention,GateDownMLP,GateRows,Reduce,DownK3iRows,Affine4Rows,Affine4Rows,Affine4Rows,Affine4Rows),
|
|
348
|
+
];
|
|
349
|
+
// Four rows: the row projections and gate/up, and the matrix down kernel.
|
|
350
|
+
this.#four=[
|
|
351
|
+
d0.stage(w,c,Delta,GateDownMLP,GateRows,Reduce,MmaDown,Affine4Rows,Affine4Rows,Affine4Rows,Affine4Rows,Affine4Rows,qk),
|
|
352
|
+
d1.stage(w,c,Delta,GateDownMLP,GateRows,Reduce,MmaDownK3i,Affine3Rows,Affine3Rows,Affine4Rows,Affine4Rows,Affine4Rows,qk),
|
|
353
|
+
d2.stage(w,c,Delta,GateDownMLP,GateRows,Reduce,MmaDownK3i,Affine3Rows,Affine3Rows,Affine4Rows,Affine4Rows,Affine4Rows,qk),
|
|
354
|
+
a3.stage(w,c,Attention,GateDownMLP,GateRows,Reduce,MmaDownK3i,Affine4Rows,Affine4Rows,Affine4Rows,Affine4Rows),
|
|
355
|
+
d4.stage(w,c,Delta,GateDownMLP,GateRows,Reduce,MmaDownK3i,Affine3Rows,Affine4Rows,Affine4Rows,Affine4Rows,Affine4Rows,qk),
|
|
356
|
+
d5.stage(w,c,Delta,GateDownMLP,GateRows,Reduce,MmaDownK3i,Affine3Rows,Affine4Rows,Affine4Rows,Affine4Rows,Affine4Rows,qk),
|
|
357
|
+
d6.stage(w,c,Delta,GateDownMLP,GateRows,Reduce,MmaDownK3i,Affine3Rows,Affine3Rows,Affine4Rows,Affine4Rows,Affine3Rows,qk),
|
|
358
|
+
a7.stage(w,c,Attention,GateDownMLP,GateRows,Reduce,MmaDownK3i,Affine4Rows,Affine4Rows,Affine4Rows,Affine4Rows),
|
|
359
|
+
d8.stage(w,c,Delta,GateDownMLP,GateRows,Reduce,MmaDownK3i,Affine3Rows,Affine3Rows,Affine4Rows,Affine4Rows,Affine3Rows,qk),
|
|
360
|
+
d9.stage(w,c,Delta,GateDownMLP,GateRows,Reduce,MmaDownK3i,Affine3Rows,Affine3Rows,Affine4Rows,Affine4Rows,Affine3Rows,qk),
|
|
361
|
+
d10.stage(w,c,Delta,GateDownMLP,GateRows,Reduce,MmaDownK3i,Affine3Rows,Affine3Rows,Affine4Rows,Affine4Rows,Affine3Rows,qk),
|
|
362
|
+
a11.stage(w,c,Attention,GateDownMLP,GateRows,Reduce,MmaDown,Affine4Rows,Affine4Rows,Affine4Rows,Affine4Rows),
|
|
363
|
+
d12.stage(w,c,Delta,GateDownMLP,GateRows,Reduce,MmaDownK3i,Affine3Rows,Affine3Rows,Affine4Rows,Affine4Rows,Affine3Rows,qk),
|
|
364
|
+
d13.stage(w,c,Delta,GateDownMLP,GateRows,Reduce,MmaDown,Affine3Rows,Affine3Rows,Affine4Rows,Affine4Rows,Affine3Rows,qk),
|
|
365
|
+
d14.stage(w,c,Delta,GateDownMLP,GateRows,Reduce,MmaDownK3i,Affine3Rows,Affine3Rows,Affine4Rows,Affine4Rows,Affine3Rows,qk),
|
|
366
|
+
a15.stage(w,c,Attention,GateDownMLP,MixedGate,Reduce,MmaDown,Affine4Rows,Affine4Rows,Affine4Rows,Affine4Rows),
|
|
367
|
+
d16.stage(w,c,Delta,GateDownMLP,GateRows,Reduce,MmaDown,Affine3Rows,Affine3Rows,Affine4Rows,Affine4Rows,Affine3Rows,qk),
|
|
368
|
+
d17.stage(w,c,Delta,GateDownMLP,GateRows,Reduce,MmaDown,Affine3Rows,Affine3Rows,Affine4Rows,Affine4Rows,Affine3Rows,qk),
|
|
369
|
+
d18.stage(w,c,Delta,GateDownMLP,GateRows,Reduce,MmaDown,Affine3Rows,Affine3Rows,Affine4Rows,Affine4Rows,Affine3Rows,qk),
|
|
370
|
+
a19.stage(w,c,Attention,GateDownMLP,GateRows,Reduce,MmaDownK3i,Affine4Rows,Affine4Rows,Affine4Rows,Affine4Rows),
|
|
371
|
+
d20.stage(w,c,Delta,GateDownMLP,GateRows,Reduce,MmaDownK3i,Affine3Rows,Affine3Rows,Affine4Rows,Affine4Rows,Affine3Rows,qk),
|
|
372
|
+
d21.stage(w,c,Delta,GateDownMLP,MixedGate,Reduce,MmaDownK3i,Affine3Rows,Affine3Rows,Affine4Rows,Affine4Rows,Affine3Rows,qk),
|
|
373
|
+
d22.stage(w,c,Delta,GateDownMLP,GateRows,Reduce,MmaDownK3i,Affine3Rows,Affine3Rows,Affine4Rows,Affine4Rows,Affine3Rows,qk),
|
|
374
|
+
a23.stage(w,c,Attention,GateDownMLP,GateRows,Reduce,MmaDownK3i,Affine4Rows,Affine4Rows,Affine4Rows,Affine4Rows),
|
|
375
|
+
d24.stage(w,c,Delta,GateDownMLP,GateRows,Reduce,MmaDownK3i,Affine3Rows,Affine3Rows,Affine4Rows,Affine4Rows,Affine3Rows,qk),
|
|
376
|
+
d25.stage(w,c,Delta,GateDownMLP,GateRows,Reduce,MmaDownK3i,Affine3Rows,Affine4Rows,Affine4Rows,Affine4Rows,Affine3Rows,qk),
|
|
377
|
+
d26.stage(w,c,Delta,GateDownMLP,GateRows,Reduce,MmaDownK3i,Affine3Rows,Affine3Rows,Affine4Rows,Affine4Rows,Affine3Rows,qk),
|
|
378
|
+
a27.stage(w,c,Attention,GateDownMLP,GateRows,Reduce,MmaDownK3i,Affine4Rows,Affine4Rows,Affine4Rows,Affine4Rows),
|
|
379
|
+
d28.stage(w,c,Delta,GateDownMLP,GateRows,Reduce,MmaDownK3i,Affine3Rows,Affine3Rows,Affine4Rows,Affine4Rows,Affine3Rows,qk),
|
|
380
|
+
d29.stage(w,c,Delta,GateDownMLP,GateRows,Reduce,MmaDown,Affine3Rows,Affine3Rows,Affine4Rows,Affine4Rows,Affine3Rows,qk),
|
|
381
|
+
d30.stage(w,c,Delta,GateDownMLP,GateRows,Reduce,MmaDownK3i,Affine3Rows,Affine3Rows,Affine4Rows,Affine4Rows,Affine3Rows,qk),
|
|
382
|
+
a31.stage(w,c,Attention,GateDownMLP,GateRows,Reduce,MmaDown,Affine4Rows,Affine4Rows,Affine4Rows,Affine4Rows),
|
|
383
|
+
d32.stage(w,c,Delta,GateDownMLP,GateRows,Reduce,MmaDown,Affine3Rows,Affine3Rows,Affine4Rows,Affine4Rows,Affine3Rows,qk),
|
|
384
|
+
d33.stage(w,c,Delta,GateDownMLP,GateRows,Reduce,MmaDown,Affine3Rows,Affine3Rows,Affine4Rows,Affine4Rows,Affine3Rows,qk),
|
|
385
|
+
d34.stage(w,c,Delta,GateDownMLP,GateRows,Reduce,MmaDownK3i,Affine3Rows,Affine3Rows,Affine4Rows,Affine4Rows,Affine3Rows,qk),
|
|
386
|
+
a35.stage(w,c,Attention,GateDownMLP,MixedGate,Reduce,MmaDownK3i,Affine4Rows,Affine4Rows,Affine4Rows,Affine4Rows),
|
|
387
|
+
d36.stage(w,c,Delta,GateDownMLP,MixedGate,Reduce,MmaDownK3i,Affine3Rows,Affine3Rows,Affine4Rows,Affine4Rows,Affine3Rows,qk),
|
|
388
|
+
d37.stage(w,c,Delta,GateDownMLP,GateRows,Reduce,MmaDownK3i,Affine3Rows,Affine3Rows,Affine4Rows,Affine4Rows,Affine3Rows,qk),
|
|
389
|
+
d38.stage(w,c,Delta,GateDownMLP,GateRows,Reduce,MmaDownK3i,Affine3Rows,Affine3Rows,Affine4Rows,Affine4Rows,Affine3Rows,qk),
|
|
390
|
+
a39.stage(w,c,Attention,GateDownMLP,GateRows,Reduce,MmaDownK3i,Affine4Rows,Affine4Rows,Affine4Rows,Affine4Rows),
|
|
391
|
+
d40.stage(w,c,Delta,GateDownMLP,GateRows,Reduce,MmaDownK3i,Affine3Rows,Affine3Rows,Affine4Rows,Affine4Rows,Affine3Rows,qk),
|
|
392
|
+
d41.stage(w,c,Delta,GateDownMLP,GateRows,Reduce,MmaDownK3i,Affine3Rows,Affine3Rows,Affine4Rows,Affine4Rows,Affine3Rows,qk),
|
|
393
|
+
d42.stage(w,c,Delta,GateDownMLP,GateRows,Reduce,MmaDownK3i,Affine3Rows,Affine3Rows,Affine4Rows,Affine4Rows,Affine3Rows,qk),
|
|
394
|
+
a43.stage(w,c,Attention,GateDownMLP,MixedGate,Reduce,MmaDownK3i,Affine4Rows,Affine4Rows,Affine4Rows,Affine4Rows),
|
|
395
|
+
d44.stage(w,c,Delta,GateDownMLP,GateRows,Reduce,MmaDownK3i,Affine3Rows,Affine3Rows,Affine4Rows,Affine4Rows,Affine3Rows,qk),
|
|
396
|
+
d45.stage(w,c,Delta,GateDownMLP,GateRows,Reduce,MmaDown,Affine3Rows,Affine3Rows,Affine4Rows,Affine4Rows,Affine3Rows,qk),
|
|
397
|
+
d46.stage(w,c,Delta,GateDownMLP,GateRows,Reduce,MmaDown,Affine3Rows,Affine3Rows,Affine4Rows,Affine4Rows,Affine3Rows,qk),
|
|
398
|
+
a47.stage(w,c,Attention,GateDownMLP,GateRows,Reduce,MmaDown,Affine4Rows,Affine4Rows,Affine4Rows,Affine4Rows),
|
|
399
|
+
d48.stage(w,c,Delta,GateDownMLP,GateRows,Reduce,MmaDown,Affine3Rows,Affine3Rows,Affine4Rows,Affine4Rows,Affine3Rows,qk),
|
|
400
|
+
d49.stage(w,c,Delta,GateDownMLP,GateRows,Reduce,MmaDown,Affine3Rows,Affine3Rows,Affine4Rows,Affine4Rows,Affine3Rows,qk),
|
|
401
|
+
d50.stage(w,c,Delta,GateDownMLP,MixedGate,Reduce,MmaDownK3i,Affine3Rows,Affine3Rows,Affine4Rows,Affine4Rows,Affine3Rows,qk),
|
|
402
|
+
a51.stage(w,c,Attention,GateDownMLP,MixedGate,Reduce,MmaDownK3i,Affine4Rows,Affine4Rows,Affine4Rows,Affine4Rows),
|
|
403
|
+
d52.stage(w,c,Delta,GateDownMLP,GateRows,Reduce,MmaDownK3i,Affine3Rows,Affine3Rows,Affine4Rows,Affine4Rows,Affine3Rows,qk),
|
|
404
|
+
d53.stage(w,c,Delta,GateDownMLP,MixedGate,Reduce,MmaDownK3i,Affine3Rows,Affine4Rows,Affine4Rows,Affine4Rows,Affine3Rows,qk),
|
|
405
|
+
d54.stage(w,c,Delta,GateDownMLP,GateRows,Reduce,MmaDown,Affine4Rows,Affine4Rows,Affine4Rows,Affine4Rows,Affine4Rows,qk),
|
|
406
|
+
a55.stage(w,c,Attention,GateDownMLP,GateRows,Reduce,MmaDown,Affine4Rows,Affine4Rows,Affine4Rows,Affine4Rows),
|
|
407
|
+
d56.stage(w,c,Delta,GateDownMLP,MixedGate,Reduce,MmaDown,Affine4Rows,Affine4Rows,Affine4Rows,Affine4Rows,Affine4Rows,qk),
|
|
408
|
+
d57.stage(w,c,Delta,GateDownMLP,GateRows,Reduce,MmaDown,Affine4Rows,Affine4Rows,Affine4Rows,Affine4Rows,Affine4Rows,qk),
|
|
409
|
+
d58.stage(w,c,Delta,GateDownMLP,GateRows,Reduce,MmaDown,Affine4Rows,Affine3Rows,Affine4Rows,Affine4Rows,Affine3Rows,qk),
|
|
410
|
+
a59.stage(w,c,Attention,GateDownMLP,GateRows,Reduce,MmaDown,Affine4Rows,Affine4Rows,Affine4Rows,Affine4Rows),
|
|
411
|
+
d60.stage(w,c,Delta,GateDownMLP,GateRows,Reduce,MmaDown,Affine4Rows,Affine4Rows,Affine4Rows,Affine4Rows,Affine3Rows,qk),
|
|
412
|
+
d61.stage(w,c,Delta,GateDownMLP,GateRows,Reduce,MmaDown,Affine4Rows,Affine4Rows,Affine4Rows,Affine4Rows,Affine4Rows,qk),
|
|
413
|
+
d62.stage(w,c,Delta,GateDownMLP,MixedGate,Reduce,MmaDownK3i,Affine4Rows,Affine4Rows,Affine4Rows,Affine4Rows,Affine3Rows,qk),
|
|
414
|
+
a63.stage(w,c,Attention,GateDownMLP,GateRows,Reduce,MmaDownK3i,Affine4Rows,Affine4Rows,Affine4Rows,Affine4Rows),
|
|
415
|
+
];
|
|
416
|
+
// Five through eight rows: matrix affine projections and Trellis gate/up/down.
|
|
417
|
+
this.#mma=[
|
|
418
|
+
d0.stage(w,c,Delta,GateDownMLP,MmaGate,Tiled,MmaDown,Affine4Mma,Affine4Mma,Affine4Mma,Affine4Mma,Affine4Mma,qk),
|
|
419
|
+
d1.stage(w,c,Delta,GateDownMLP,MmaGate,Tiled,MmaDownK3i,Affine3Mma,Affine3Mma,Affine4Mma,Affine4Mma,Affine4Mma,qk),
|
|
420
|
+
d2.stage(w,c,Delta,GateDownMLP,MmaGate,Tiled,MmaDownK3i,Affine3Mma,Affine3Mma,Affine4Mma,Affine4Mma,Affine4Mma,qk),
|
|
421
|
+
a3.stage(w,c,Attention,GateDownMLP,MmaGate,Tiled,MmaDownK3i,Affine4Mma,Affine4Mma,Affine4Mma,Affine4Mma),
|
|
422
|
+
d4.stage(w,c,Delta,GateDownMLP,MmaGate,Tiled,MmaDownK3i,Affine3Mma,Affine4Mma,Affine4Mma,Affine4Mma,Affine4Mma,qk),
|
|
423
|
+
d5.stage(w,c,Delta,GateDownMLP,MmaGate,Tiled,MmaDownK3i,Affine3Mma,Affine4Mma,Affine4Mma,Affine4Mma,Affine4Mma,qk),
|
|
424
|
+
d6.stage(w,c,Delta,GateDownMLP,MmaGate,Tiled,MmaDownK3i,Affine3Mma,Affine3Mma,Affine4Mma,Affine4Mma,Affine3Mma,qk),
|
|
425
|
+
a7.stage(w,c,Attention,GateDownMLP,MmaGate,Tiled,MmaDownK3i,Affine4Mma,Affine4Mma,Affine4Mma,Affine4Mma),
|
|
426
|
+
d8.stage(w,c,Delta,GateDownMLP,MmaGate,Tiled,MmaDownK3i,Affine3Mma,Affine3Mma,Affine4Mma,Affine4Mma,Affine3Mma,qk),
|
|
427
|
+
d9.stage(w,c,Delta,GateDownMLP,MmaGate,Tiled,MmaDownK3i,Affine3Mma,Affine3Mma,Affine4Mma,Affine4Mma,Affine3Mma,qk),
|
|
428
|
+
d10.stage(w,c,Delta,GateDownMLP,MmaGate,Tiled,MmaDownK3i,Affine3Mma,Affine3Mma,Affine4Mma,Affine4Mma,Affine3Mma,qk),
|
|
429
|
+
a11.stage(w,c,Attention,GateDownMLP,MmaGate,Tiled,MmaDown,Affine4Mma,Affine4Mma,Affine4Mma,Affine4Mma),
|
|
430
|
+
d12.stage(w,c,Delta,GateDownMLP,MmaGate,Tiled,MmaDownK3i,Affine3Mma,Affine3Mma,Affine4Mma,Affine4Mma,Affine3Mma,qk),
|
|
431
|
+
d13.stage(w,c,Delta,GateDownMLP,MmaGate,Tiled,MmaDown,Affine3Mma,Affine3Mma,Affine4Mma,Affine4Mma,Affine3Mma,qk),
|
|
432
|
+
d14.stage(w,c,Delta,GateDownMLP,MmaGate,Tiled,MmaDownK3i,Affine3Mma,Affine3Mma,Affine4Mma,Affine4Mma,Affine3Mma,qk),
|
|
433
|
+
a15.stage(w,c,Attention,GateDownMLP,MmaGate,Tiled,MmaDown,Affine4Mma,Affine4Mma,Affine4Mma,Affine4Mma),
|
|
434
|
+
d16.stage(w,c,Delta,GateDownMLP,MmaGate,Tiled,MmaDown,Affine3Mma,Affine3Mma,Affine4Mma,Affine4Mma,Affine3Mma,qk),
|
|
435
|
+
d17.stage(w,c,Delta,GateDownMLP,MmaGate,Tiled,MmaDown,Affine3Mma,Affine3Mma,Affine4Mma,Affine4Mma,Affine3Mma,qk),
|
|
436
|
+
d18.stage(w,c,Delta,GateDownMLP,MmaGate,Tiled,MmaDown,Affine3Mma,Affine3Mma,Affine4Mma,Affine4Mma,Affine3Mma,qk),
|
|
437
|
+
a19.stage(w,c,Attention,GateDownMLP,MmaGate,Tiled,MmaDownK3i,Affine4Mma,Affine4Mma,Affine4Mma,Affine4Mma),
|
|
438
|
+
d20.stage(w,c,Delta,GateDownMLP,MmaGate,Tiled,MmaDownK3i,Affine3Mma,Affine3Mma,Affine4Mma,Affine4Mma,Affine3Mma,qk),
|
|
439
|
+
d21.stage(w,c,Delta,GateDownMLP,MmaGate,Tiled,MmaDownK3i,Affine3Mma,Affine3Mma,Affine4Mma,Affine4Mma,Affine3Mma,qk),
|
|
440
|
+
d22.stage(w,c,Delta,GateDownMLP,MmaGate,Tiled,MmaDownK3i,Affine3Mma,Affine3Mma,Affine4Mma,Affine4Mma,Affine3Mma,qk),
|
|
441
|
+
a23.stage(w,c,Attention,GateDownMLP,MmaGate,Tiled,MmaDownK3i,Affine4Mma,Affine4Mma,Affine4Mma,Affine4Mma),
|
|
442
|
+
d24.stage(w,c,Delta,GateDownMLP,MmaGate,Tiled,MmaDownK3i,Affine3Mma,Affine3Mma,Affine4Mma,Affine4Mma,Affine3Mma,qk),
|
|
443
|
+
d25.stage(w,c,Delta,GateDownMLP,MmaGate,Tiled,MmaDownK3i,Affine3Mma,Affine4Mma,Affine4Mma,Affine4Mma,Affine3Mma,qk),
|
|
444
|
+
d26.stage(w,c,Delta,GateDownMLP,MmaGate,Tiled,MmaDownK3i,Affine3Mma,Affine3Mma,Affine4Mma,Affine4Mma,Affine3Mma,qk),
|
|
445
|
+
a27.stage(w,c,Attention,GateDownMLP,MmaGate,Tiled,MmaDownK3i,Affine4Mma,Affine4Mma,Affine4Mma,Affine4Mma),
|
|
446
|
+
d28.stage(w,c,Delta,GateDownMLP,MmaGate,Tiled,MmaDownK3i,Affine3Mma,Affine3Mma,Affine4Mma,Affine4Mma,Affine3Mma,qk),
|
|
447
|
+
d29.stage(w,c,Delta,GateDownMLP,MmaGate,Tiled,MmaDown,Affine3Mma,Affine3Mma,Affine4Mma,Affine4Mma,Affine3Mma,qk),
|
|
448
|
+
d30.stage(w,c,Delta,GateDownMLP,MmaGate,Tiled,MmaDownK3i,Affine3Mma,Affine3Mma,Affine4Mma,Affine4Mma,Affine3Mma,qk),
|
|
449
|
+
a31.stage(w,c,Attention,GateDownMLP,MmaGate,Tiled,MmaDown,Affine4Mma,Affine4Mma,Affine4Mma,Affine4Mma),
|
|
450
|
+
d32.stage(w,c,Delta,GateDownMLP,MmaGate,Tiled,MmaDown,Affine3Mma,Affine3Mma,Affine4Mma,Affine4Mma,Affine3Mma,qk),
|
|
451
|
+
d33.stage(w,c,Delta,GateDownMLP,MmaGate,Tiled,MmaDown,Affine3Mma,Affine3Mma,Affine4Mma,Affine4Mma,Affine3Mma,qk),
|
|
452
|
+
d34.stage(w,c,Delta,GateDownMLP,MmaGate,Tiled,MmaDownK3i,Affine3Mma,Affine3Mma,Affine4Mma,Affine4Mma,Affine3Mma,qk),
|
|
453
|
+
a35.stage(w,c,Attention,GateDownMLP,MmaGate,Tiled,MmaDownK3i,Affine4Mma,Affine4Mma,Affine4Mma,Affine4Mma),
|
|
454
|
+
d36.stage(w,c,Delta,GateDownMLP,MmaGate,Tiled,MmaDownK3i,Affine3Mma,Affine3Mma,Affine4Mma,Affine4Mma,Affine3Mma,qk),
|
|
455
|
+
d37.stage(w,c,Delta,GateDownMLP,MmaGate,Tiled,MmaDownK3i,Affine3Mma,Affine3Mma,Affine4Mma,Affine4Mma,Affine3Mma,qk),
|
|
456
|
+
d38.stage(w,c,Delta,GateDownMLP,MmaGate,Tiled,MmaDownK3i,Affine3Mma,Affine3Mma,Affine4Mma,Affine4Mma,Affine3Mma,qk),
|
|
457
|
+
a39.stage(w,c,Attention,GateDownMLP,MmaGate,Tiled,MmaDownK3i,Affine4Mma,Affine4Mma,Affine4Mma,Affine4Mma),
|
|
458
|
+
d40.stage(w,c,Delta,GateDownMLP,MmaGate,Tiled,MmaDownK3i,Affine3Mma,Affine3Mma,Affine4Mma,Affine4Mma,Affine3Mma,qk),
|
|
459
|
+
d41.stage(w,c,Delta,GateDownMLP,MmaGate,Tiled,MmaDownK3i,Affine3Mma,Affine3Mma,Affine4Mma,Affine4Mma,Affine3Mma,qk),
|
|
460
|
+
d42.stage(w,c,Delta,GateDownMLP,MmaGate,Tiled,MmaDownK3i,Affine3Mma,Affine3Mma,Affine4Mma,Affine4Mma,Affine3Mma,qk),
|
|
461
|
+
a43.stage(w,c,Attention,GateDownMLP,MmaGate,Tiled,MmaDownK3i,Affine4Mma,Affine4Mma,Affine4Mma,Affine4Mma),
|
|
462
|
+
d44.stage(w,c,Delta,GateDownMLP,MmaGate,Tiled,MmaDownK3i,Affine3Mma,Affine3Mma,Affine4Mma,Affine4Mma,Affine3Mma,qk),
|
|
463
|
+
d45.stage(w,c,Delta,GateDownMLP,MmaGate,Tiled,MmaDown,Affine3Mma,Affine3Mma,Affine4Mma,Affine4Mma,Affine3Mma,qk),
|
|
464
|
+
d46.stage(w,c,Delta,GateDownMLP,MmaGate,Tiled,MmaDown,Affine3Mma,Affine3Mma,Affine4Mma,Affine4Mma,Affine3Mma,qk),
|
|
465
|
+
a47.stage(w,c,Attention,GateDownMLP,MmaGate,Tiled,MmaDown,Affine4Mma,Affine4Mma,Affine4Mma,Affine4Mma),
|
|
466
|
+
d48.stage(w,c,Delta,GateDownMLP,MmaGate,Tiled,MmaDown,Affine3Mma,Affine3Mma,Affine4Mma,Affine4Mma,Affine3Mma,qk),
|
|
467
|
+
d49.stage(w,c,Delta,GateDownMLP,MmaGate,Tiled,MmaDown,Affine3Mma,Affine3Mma,Affine4Mma,Affine4Mma,Affine3Mma,qk),
|
|
468
|
+
d50.stage(w,c,Delta,GateDownMLP,MmaGate,Tiled,MmaDownK3i,Affine3Mma,Affine3Mma,Affine4Mma,Affine4Mma,Affine3Mma,qk),
|
|
469
|
+
a51.stage(w,c,Attention,GateDownMLP,MmaGate,Tiled,MmaDownK3i,Affine4Mma,Affine4Mma,Affine4Mma,Affine4Mma),
|
|
470
|
+
d52.stage(w,c,Delta,GateDownMLP,MmaGate,Tiled,MmaDownK3i,Affine3Mma,Affine3Mma,Affine4Mma,Affine4Mma,Affine3Mma,qk),
|
|
471
|
+
d53.stage(w,c,Delta,GateDownMLP,MmaGate,Tiled,MmaDownK3i,Affine3Mma,Affine4Mma,Affine4Mma,Affine4Mma,Affine3Mma,qk),
|
|
472
|
+
d54.stage(w,c,Delta,GateDownMLP,MmaGate,Tiled,MmaDown,Affine4Mma,Affine4Mma,Affine4Mma,Affine4Mma,Affine4Mma,qk),
|
|
473
|
+
a55.stage(w,c,Attention,GateDownMLP,MmaGate,Tiled,MmaDown,Affine4Mma,Affine4Mma,Affine4Mma,Affine4Mma),
|
|
474
|
+
d56.stage(w,c,Delta,GateDownMLP,MmaGate,Tiled,MmaDown,Affine4Mma,Affine4Mma,Affine4Mma,Affine4Mma,Affine4Mma,qk),
|
|
475
|
+
d57.stage(w,c,Delta,GateDownMLP,MmaGate,Tiled,MmaDown,Affine4Mma,Affine4Mma,Affine4Mma,Affine4Mma,Affine4Mma,qk),
|
|
476
|
+
d58.stage(w,c,Delta,GateDownMLP,MmaGate,Tiled,MmaDown,Affine4Mma,Affine3Mma,Affine4Mma,Affine4Mma,Affine3Mma,qk),
|
|
477
|
+
a59.stage(w,c,Attention,GateDownMLP,MmaGate,Tiled,MmaDown,Affine4Mma,Affine4Mma,Affine4Mma,Affine4Mma),
|
|
478
|
+
d60.stage(w,c,Delta,GateDownMLP,MmaGate,Tiled,MmaDown,Affine4Mma,Affine4Mma,Affine4Mma,Affine4Mma,Affine3Mma,qk),
|
|
479
|
+
d61.stage(w,c,Delta,GateDownMLP,MmaGate,Tiled,MmaDown,Affine4Mma,Affine4Mma,Affine4Mma,Affine4Mma,Affine4Mma,qk),
|
|
480
|
+
d62.stage(w,c,Delta,GateDownMLP,MmaGate,Tiled,MmaDownK3i,Affine4Mma,Affine4Mma,Affine4Mma,Affine4Mma,Affine3Mma,qk),
|
|
481
|
+
a63.stage(w,c,Attention,GateDownMLP,MmaGate,Tiled,MmaDownK3i,Affine4Mma,Affine4Mma,Affine4Mma,Affine4Mma),
|
|
482
|
+
];
|
|
483
|
+
// Native GemvWide gate/up, followed by the evaluated expanded down.
|
|
484
|
+
this.#wide=[
|
|
485
|
+
d0.prompt(w,c,Delta,TrellisUnfusedMLP,Wide,ExpandDown,Plain,qk),d1.prompt(w,c,Delta,TrellisUnfusedMLP,Wide,ExpandDown,Plain,qk),d2.prompt(w,c,Delta,TrellisUnfusedMLP,Wide,ExpandDown,Plain,qk),a3.prompt(w,c,Attention,TrellisUnfusedMLP,Wide,ExpandDown,Plain),
|
|
486
|
+
d4.prompt(w,c,Delta,TrellisUnfusedMLP,Wide,ExpandDown,Plain,qk),d5.prompt(w,c,Delta,TrellisUnfusedMLP,Wide,ExpandDown,Plain,qk),d6.prompt(w,c,Delta,TrellisUnfusedMLP,Wide,ExpandDown,Plain,qk),a7.prompt(w,c,Attention,TrellisUnfusedMLP,Wide,ExpandDown,Plain),
|
|
487
|
+
d8.prompt(w,c,Delta,TrellisUnfusedMLP,Wide,ExpandDown,Plain,qk),d9.prompt(w,c,Delta,TrellisUnfusedMLP,Wide,ExpandDown,Plain,qk),d10.prompt(w,c,Delta,TrellisUnfusedMLP,Wide,ExpandDown,Plain,qk),a11.prompt(w,c,Attention,TrellisUnfusedMLP,Wide,ExpandDown,Plain),
|
|
488
|
+
d12.prompt(w,c,Delta,TrellisUnfusedMLP,Wide,ExpandDown,Plain,qk),d13.prompt(w,c,Delta,TrellisUnfusedMLP,Wide,ExpandDown,Plain,qk),d14.prompt(w,c,Delta,TrellisUnfusedMLP,Wide,ExpandDown,Plain,qk),a15.prompt(w,c,Attention,TrellisUnfusedMLP,Wide,ExpandDown,Plain),
|
|
489
|
+
d16.prompt(w,c,Delta,TrellisUnfusedMLP,Wide,ExpandDown,Plain,qk),d17.prompt(w,c,Delta,TrellisUnfusedMLP,Wide,ExpandDown,Plain,qk),d18.prompt(w,c,Delta,TrellisUnfusedMLP,Wide,ExpandDown,Plain,qk),a19.prompt(w,c,Attention,TrellisUnfusedMLP,Wide,ExpandDown,Plain),
|
|
490
|
+
d20.prompt(w,c,Delta,TrellisUnfusedMLP,Wide,ExpandDown,Plain,qk),d21.prompt(w,c,Delta,TrellisUnfusedMLP,Wide,ExpandDown,Plain,qk),d22.prompt(w,c,Delta,TrellisUnfusedMLP,Wide,ExpandDown,Plain,qk),a23.prompt(w,c,Attention,TrellisUnfusedMLP,Wide,ExpandDown,Plain),
|
|
491
|
+
d24.prompt(w,c,Delta,TrellisUnfusedMLP,Wide,ExpandDown,Plain,qk),d25.prompt(w,c,Delta,TrellisUnfusedMLP,Wide,ExpandDown,Plain,qk),d26.prompt(w,c,Delta,TrellisUnfusedMLP,Wide,ExpandDown,Plain,qk),a27.prompt(w,c,Attention,TrellisUnfusedMLP,Wide,ExpandDown,Plain),
|
|
492
|
+
d28.prompt(w,c,Delta,TrellisUnfusedMLP,Wide,ExpandDown,Plain,qk),d29.prompt(w,c,Delta,TrellisUnfusedMLP,Wide,ExpandDown,Plain,qk),d30.prompt(w,c,Delta,TrellisUnfusedMLP,Wide,ExpandDown,Plain,qk),a31.prompt(w,c,Attention,TrellisUnfusedMLP,Wide,ExpandDown,Plain),
|
|
493
|
+
d32.prompt(w,c,Delta,TrellisUnfusedMLP,Wide,ExpandDown,Plain,qk),d33.prompt(w,c,Delta,TrellisUnfusedMLP,Wide,ExpandDown,Plain,qk),d34.prompt(w,c,Delta,TrellisUnfusedMLP,Wide,ExpandDown,Plain,qk),a35.prompt(w,c,Attention,TrellisUnfusedMLP,Wide,ExpandDown,Plain),
|
|
494
|
+
d36.prompt(w,c,Delta,TrellisUnfusedMLP,Wide,ExpandDown,Plain,qk),d37.prompt(w,c,Delta,TrellisUnfusedMLP,Wide,ExpandDown,Plain,qk),d38.prompt(w,c,Delta,TrellisUnfusedMLP,Wide,ExpandDown,Plain,qk),a39.prompt(w,c,Attention,TrellisUnfusedMLP,Wide,ExpandDown,Plain),
|
|
495
|
+
d40.prompt(w,c,Delta,TrellisUnfusedMLP,Wide,ExpandDown,Plain,qk),d41.prompt(w,c,Delta,TrellisUnfusedMLP,Wide,ExpandDown,Plain,qk),d42.prompt(w,c,Delta,TrellisUnfusedMLP,Wide,ExpandDown,Plain,qk),a43.prompt(w,c,Attention,TrellisUnfusedMLP,Wide,ExpandDown,Plain),
|
|
496
|
+
d44.prompt(w,c,Delta,TrellisUnfusedMLP,Wide,ExpandDown,Plain,qk),d45.prompt(w,c,Delta,TrellisUnfusedMLP,Wide,ExpandDown,Plain,qk),d46.prompt(w,c,Delta,TrellisUnfusedMLP,Wide,ExpandDown,Plain,qk),a47.prompt(w,c,Attention,TrellisUnfusedMLP,Wide,ExpandDown,Plain),
|
|
497
|
+
d48.prompt(w,c,Delta,TrellisUnfusedMLP,Wide,ExpandDown,Plain,qk),d49.prompt(w,c,Delta,TrellisUnfusedMLP,Wide,ExpandDown,Plain,qk),d50.prompt(w,c,Delta,TrellisUnfusedMLP,Wide,ExpandDown,Plain,qk),a51.prompt(w,c,Attention,TrellisUnfusedMLP,Wide,ExpandDown,Plain),
|
|
498
|
+
d52.prompt(w,c,Delta,TrellisUnfusedMLP,Wide,ExpandDown,Plain,qk),d53.prompt(w,c,Delta,TrellisUnfusedMLP,Wide,ExpandDown,Plain,qk),d54.prompt(w,c,Delta,TrellisUnfusedMLP,Wide,ExpandDown,Plain,qk),a55.prompt(w,c,Attention,TrellisUnfusedMLP,Wide,ExpandDown,Plain),
|
|
499
|
+
d56.prompt(w,c,Delta,TrellisUnfusedMLP,Wide,ExpandDown,Plain,qk),d57.prompt(w,c,Delta,TrellisUnfusedMLP,Wide,ExpandDown,Plain,qk),d58.prompt(w,c,Delta,TrellisUnfusedMLP,Wide,ExpandDown,Plain,qk),a59.prompt(w,c,Attention,TrellisUnfusedMLP,Wide,ExpandDown,Plain),
|
|
500
|
+
d60.prompt(w,c,Delta,TrellisUnfusedMLP,Wide,ExpandDown,Plain,qk),d61.prompt(w,c,Delta,TrellisUnfusedMLP,Wide,ExpandDown,Plain,qk),d62.prompt(w,c,Delta,TrellisUnfusedMLP,Wide,ExpandDown,Plain,qk),a63.prompt(w,c,Attention,TrellisUnfusedMLP,Wide,ExpandDown,Plain),
|
|
501
|
+
];
|
|
502
|
+
this.#tiled=[
|
|
503
|
+
d0.prompt(w,c,Delta,TrellisUnfusedMLP,Tiled,ExpandDown,Plain,qk),d1.prompt(w,c,Delta,TrellisUnfusedMLP,Tiled,ExpandDown,Plain,qk),d2.prompt(w,c,Delta,TrellisUnfusedMLP,Tiled,ExpandDown,Plain,qk),a3.prompt(w,c,Attention,TrellisUnfusedMLP,Tiled,ExpandDown,Plain),
|
|
504
|
+
d4.prompt(w,c,Delta,TrellisUnfusedMLP,Tiled,ExpandDown,Plain,qk),d5.prompt(w,c,Delta,TrellisUnfusedMLP,Tiled,ExpandDown,Plain,qk),d6.prompt(w,c,Delta,TrellisUnfusedMLP,Tiled,ExpandDown,Plain,qk),a7.prompt(w,c,Attention,TrellisUnfusedMLP,Tiled,ExpandDown,Plain),
|
|
505
|
+
d8.prompt(w,c,Delta,TrellisUnfusedMLP,Tiled,ExpandDown,Plain,qk),d9.prompt(w,c,Delta,TrellisUnfusedMLP,Tiled,ExpandDown,Plain,qk),d10.prompt(w,c,Delta,TrellisUnfusedMLP,Tiled,ExpandDown,Plain,qk),a11.prompt(w,c,Attention,TrellisUnfusedMLP,Tiled,ExpandDown,Plain),
|
|
506
|
+
d12.prompt(w,c,Delta,TrellisUnfusedMLP,Tiled,ExpandDown,Plain,qk),d13.prompt(w,c,Delta,TrellisUnfusedMLP,Tiled,ExpandDown,Plain,qk),d14.prompt(w,c,Delta,TrellisUnfusedMLP,Tiled,ExpandDown,Plain,qk),a15.prompt(w,c,Attention,TrellisUnfusedMLP,Tiled,ExpandDown,Plain),
|
|
507
|
+
d16.prompt(w,c,Delta,TrellisUnfusedMLP,Tiled,ExpandDown,Plain,qk),d17.prompt(w,c,Delta,TrellisUnfusedMLP,Tiled,ExpandDown,Plain,qk),d18.prompt(w,c,Delta,TrellisUnfusedMLP,Tiled,ExpandDown,Plain,qk),a19.prompt(w,c,Attention,TrellisUnfusedMLP,Tiled,ExpandDown,Plain),
|
|
508
|
+
d20.prompt(w,c,Delta,TrellisUnfusedMLP,Tiled,ExpandDown,Plain,qk),d21.prompt(w,c,Delta,TrellisUnfusedMLP,Tiled,ExpandDown,Plain,qk),d22.prompt(w,c,Delta,TrellisUnfusedMLP,Tiled,ExpandDown,Plain,qk),a23.prompt(w,c,Attention,TrellisUnfusedMLP,Tiled,ExpandDown,Plain),
|
|
509
|
+
d24.prompt(w,c,Delta,TrellisUnfusedMLP,Tiled,ExpandDown,Plain,qk),d25.prompt(w,c,Delta,TrellisUnfusedMLP,Tiled,ExpandDown,Plain,qk),d26.prompt(w,c,Delta,TrellisUnfusedMLP,Tiled,ExpandDown,Plain,qk),a27.prompt(w,c,Attention,TrellisUnfusedMLP,Tiled,ExpandDown,Plain),
|
|
510
|
+
d28.prompt(w,c,Delta,TrellisUnfusedMLP,Tiled,ExpandDown,Plain,qk),d29.prompt(w,c,Delta,TrellisUnfusedMLP,Tiled,ExpandDown,Plain,qk),d30.prompt(w,c,Delta,TrellisUnfusedMLP,Tiled,ExpandDown,Plain,qk),a31.prompt(w,c,Attention,TrellisUnfusedMLP,Tiled,ExpandDown,Plain),
|
|
511
|
+
d32.prompt(w,c,Delta,TrellisUnfusedMLP,Tiled,ExpandDown,Plain,qk),d33.prompt(w,c,Delta,TrellisUnfusedMLP,Tiled,ExpandDown,Plain,qk),d34.prompt(w,c,Delta,TrellisUnfusedMLP,Tiled,ExpandDown,Plain,qk),a35.prompt(w,c,Attention,TrellisUnfusedMLP,Tiled,ExpandDown,Plain),
|
|
512
|
+
d36.prompt(w,c,Delta,TrellisUnfusedMLP,Tiled,ExpandDown,Plain,qk),d37.prompt(w,c,Delta,TrellisUnfusedMLP,Tiled,ExpandDown,Plain,qk),d38.prompt(w,c,Delta,TrellisUnfusedMLP,Tiled,ExpandDown,Plain,qk),a39.prompt(w,c,Attention,TrellisUnfusedMLP,Tiled,ExpandDown,Plain),
|
|
513
|
+
d40.prompt(w,c,Delta,TrellisUnfusedMLP,Tiled,ExpandDown,Plain,qk),d41.prompt(w,c,Delta,TrellisUnfusedMLP,Tiled,ExpandDown,Plain,qk),d42.prompt(w,c,Delta,TrellisUnfusedMLP,Tiled,ExpandDown,Plain,qk),a43.prompt(w,c,Attention,TrellisUnfusedMLP,Tiled,ExpandDown,Plain),
|
|
514
|
+
d44.prompt(w,c,Delta,TrellisUnfusedMLP,Tiled,ExpandDown,Plain,qk),d45.prompt(w,c,Delta,TrellisUnfusedMLP,Tiled,ExpandDown,Plain,qk),d46.prompt(w,c,Delta,TrellisUnfusedMLP,Tiled,ExpandDown,Plain,qk),a47.prompt(w,c,Attention,TrellisUnfusedMLP,Tiled,ExpandDown,Plain),
|
|
515
|
+
d48.prompt(w,c,Delta,TrellisUnfusedMLP,Tiled,ExpandDown,Plain,qk),d49.prompt(w,c,Delta,TrellisUnfusedMLP,Tiled,ExpandDown,Plain,qk),d50.prompt(w,c,Delta,TrellisUnfusedMLP,Tiled,ExpandDown,Plain,qk),a51.prompt(w,c,Attention,TrellisUnfusedMLP,Tiled,ExpandDown,Plain),
|
|
516
|
+
d52.prompt(w,c,Delta,TrellisUnfusedMLP,Tiled,ExpandDown,Plain,qk),d53.prompt(w,c,Delta,TrellisUnfusedMLP,Tiled,ExpandDown,Plain,qk),d54.prompt(w,c,Delta,TrellisUnfusedMLP,Tiled,ExpandDown,Plain,qk),a55.prompt(w,c,Attention,TrellisUnfusedMLP,Tiled,ExpandDown,Plain),
|
|
517
|
+
d56.prompt(w,c,Delta,TrellisUnfusedMLP,Tiled,ExpandDown,Plain,qk),d57.prompt(w,c,Delta,TrellisUnfusedMLP,Tiled,ExpandDown,Plain,qk),d58.prompt(w,c,Delta,TrellisUnfusedMLP,Tiled,ExpandDown,Plain,qk),a59.prompt(w,c,Attention,TrellisUnfusedMLP,Tiled,ExpandDown,Plain),
|
|
518
|
+
d60.prompt(w,c,Delta,TrellisUnfusedMLP,Tiled,ExpandDown,Plain,qk),d61.prompt(w,c,Delta,TrellisUnfusedMLP,Tiled,ExpandDown,Plain,qk),d62.prompt(w,c,Delta,TrellisUnfusedMLP,Tiled,ExpandDown,Plain,qk),a63.prompt(w,c,Attention,TrellisUnfusedMLP,Tiled,ExpandDown,Plain),
|
|
519
|
+
];
|
|
520
|
+
this.#expanded=[
|
|
521
|
+
d0.prompt(w,c,Delta,TrellisUnfusedMLP,ExpandGate,ExpandDown,Plain,qk),d1.prompt(w,c,Delta,TrellisUnfusedMLP,ExpandGate,ExpandDown,Plain,qk),d2.prompt(w,c,Delta,TrellisUnfusedMLP,ExpandGate,ExpandDown,Plain,qk),a3.prompt(w,c,Attention,TrellisUnfusedMLP,ExpandGate,ExpandDown,Plain),
|
|
522
|
+
d4.prompt(w,c,Delta,TrellisUnfusedMLP,ExpandGate,ExpandDown,Plain,qk),d5.prompt(w,c,Delta,TrellisUnfusedMLP,ExpandGate,ExpandDown,Plain,qk),d6.prompt(w,c,Delta,TrellisUnfusedMLP,ExpandGate,ExpandDown,Plain,qk),a7.prompt(w,c,Attention,TrellisUnfusedMLP,ExpandGate,ExpandDown,Plain),
|
|
523
|
+
d8.prompt(w,c,Delta,TrellisUnfusedMLP,ExpandGate,ExpandDown,Plain,qk),d9.prompt(w,c,Delta,TrellisUnfusedMLP,ExpandGate,ExpandDown,Plain,qk),d10.prompt(w,c,Delta,TrellisUnfusedMLP,ExpandGate,ExpandDown,Plain,qk),a11.prompt(w,c,Attention,TrellisUnfusedMLP,ExpandGate,ExpandDown,Plain),
|
|
524
|
+
d12.prompt(w,c,Delta,TrellisUnfusedMLP,ExpandGate,ExpandDown,Plain,qk),d13.prompt(w,c,Delta,TrellisUnfusedMLP,ExpandGate,ExpandDown,Plain,qk),d14.prompt(w,c,Delta,TrellisUnfusedMLP,ExpandGate,ExpandDown,Plain,qk),a15.prompt(w,c,Attention,TrellisUnfusedMLP,ExpandGate,ExpandDown,Plain),
|
|
525
|
+
d16.prompt(w,c,Delta,TrellisUnfusedMLP,ExpandGate,ExpandDown,Plain,qk),d17.prompt(w,c,Delta,TrellisUnfusedMLP,ExpandGate,ExpandDown,Plain,qk),d18.prompt(w,c,Delta,TrellisUnfusedMLP,ExpandGate,ExpandDown,Plain,qk),a19.prompt(w,c,Attention,TrellisUnfusedMLP,ExpandGate,ExpandDown,Plain),
|
|
526
|
+
d20.prompt(w,c,Delta,TrellisUnfusedMLP,ExpandGate,ExpandDown,Plain,qk),d21.prompt(w,c,Delta,TrellisUnfusedMLP,ExpandGate,ExpandDown,Plain,qk),d22.prompt(w,c,Delta,TrellisUnfusedMLP,ExpandGate,ExpandDown,Plain,qk),a23.prompt(w,c,Attention,TrellisUnfusedMLP,ExpandGate,ExpandDown,Plain),
|
|
527
|
+
d24.prompt(w,c,Delta,TrellisUnfusedMLP,ExpandGate,ExpandDown,Plain,qk),d25.prompt(w,c,Delta,TrellisUnfusedMLP,ExpandGate,ExpandDown,Plain,qk),d26.prompt(w,c,Delta,TrellisUnfusedMLP,ExpandGate,ExpandDown,Plain,qk),a27.prompt(w,c,Attention,TrellisUnfusedMLP,ExpandGate,ExpandDown,Plain),
|
|
528
|
+
d28.prompt(w,c,Delta,TrellisUnfusedMLP,ExpandGate,ExpandDown,Plain,qk),d29.prompt(w,c,Delta,TrellisUnfusedMLP,ExpandGate,ExpandDown,Plain,qk),d30.prompt(w,c,Delta,TrellisUnfusedMLP,ExpandGate,ExpandDown,Plain,qk),a31.prompt(w,c,Attention,TrellisUnfusedMLP,ExpandGate,ExpandDown,Plain),
|
|
529
|
+
d32.prompt(w,c,Delta,TrellisUnfusedMLP,ExpandGate,ExpandDown,Plain,qk),d33.prompt(w,c,Delta,TrellisUnfusedMLP,ExpandGate,ExpandDown,Plain,qk),d34.prompt(w,c,Delta,TrellisUnfusedMLP,ExpandGate,ExpandDown,Plain,qk),a35.prompt(w,c,Attention,TrellisUnfusedMLP,ExpandGate,ExpandDown,Plain),
|
|
530
|
+
d36.prompt(w,c,Delta,TrellisUnfusedMLP,ExpandGate,ExpandDown,Plain,qk),d37.prompt(w,c,Delta,TrellisUnfusedMLP,ExpandGate,ExpandDown,Plain,qk),d38.prompt(w,c,Delta,TrellisUnfusedMLP,ExpandGate,ExpandDown,Plain,qk),a39.prompt(w,c,Attention,TrellisUnfusedMLP,ExpandGate,ExpandDown,Plain),
|
|
531
|
+
d40.prompt(w,c,Delta,TrellisUnfusedMLP,ExpandGate,ExpandDown,Plain,qk),d41.prompt(w,c,Delta,TrellisUnfusedMLP,ExpandGate,ExpandDown,Plain,qk),d42.prompt(w,c,Delta,TrellisUnfusedMLP,ExpandGate,ExpandDown,Plain,qk),a43.prompt(w,c,Attention,TrellisUnfusedMLP,ExpandGate,ExpandDown,Plain),
|
|
532
|
+
d44.prompt(w,c,Delta,TrellisUnfusedMLP,ExpandGate,ExpandDown,Plain,qk),d45.prompt(w,c,Delta,TrellisUnfusedMLP,ExpandGate,ExpandDown,Plain,qk),d46.prompt(w,c,Delta,TrellisUnfusedMLP,ExpandGate,ExpandDown,Plain,qk),a47.prompt(w,c,Attention,TrellisUnfusedMLP,ExpandGate,ExpandDown,Plain),
|
|
533
|
+
d48.prompt(w,c,Delta,TrellisUnfusedMLP,ExpandGate,ExpandDown,Plain,qk),d49.prompt(w,c,Delta,TrellisUnfusedMLP,ExpandGate,ExpandDown,Plain,qk),d50.prompt(w,c,Delta,TrellisUnfusedMLP,ExpandGate,ExpandDown,Plain,qk),a51.prompt(w,c,Attention,TrellisUnfusedMLP,ExpandGate,ExpandDown,Plain),
|
|
534
|
+
d52.prompt(w,c,Delta,TrellisUnfusedMLP,ExpandGate,ExpandDown,Plain,qk),d53.prompt(w,c,Delta,TrellisUnfusedMLP,ExpandGate,ExpandDown,Plain,qk),d54.prompt(w,c,Delta,TrellisUnfusedMLP,ExpandGate,ExpandDown,Plain,qk),a55.prompt(w,c,Attention,TrellisUnfusedMLP,ExpandGate,ExpandDown,Plain),
|
|
535
|
+
d56.prompt(w,c,Delta,TrellisUnfusedMLP,ExpandGate,ExpandDown,Plain,qk),d57.prompt(w,c,Delta,TrellisUnfusedMLP,ExpandGate,ExpandDown,Plain,qk),d58.prompt(w,c,Delta,TrellisUnfusedMLP,ExpandGate,ExpandDown,Plain,qk),a59.prompt(w,c,Attention,TrellisUnfusedMLP,ExpandGate,ExpandDown,Plain),
|
|
536
|
+
d60.prompt(w,c,Delta,TrellisUnfusedMLP,ExpandGate,ExpandDown,Plain,qk),d61.prompt(w,c,Delta,TrellisUnfusedMLP,ExpandGate,ExpandDown,Plain,qk),d62.prompt(w,c,Delta,TrellisUnfusedMLP,ExpandGate,ExpandDown,Plain,qk),a63.prompt(w,c,Attention,TrellisUnfusedMLP,ExpandGate,ExpandDown,Plain),
|
|
537
|
+
];
|
|
538
|
+
// Largest programs first: later ranges reuse their buffers. Each layer's
|
|
539
|
+
// MLP and each eligible affine projection names its ANE split explicitly.
|
|
540
|
+
this.#ane2048=[
|
|
541
|
+
d0.aneStage(w,c,Delta,AneBoundMlp,AneMlp,AneAffineSplit,Plain,ANE_FRACTION,2048,shares,qk),d1.aneStage(w,c,Delta,AneBoundMlp,AneMlpK3i,AneAffineSplit,Plain,ANE_FRACTION,2048,shares,qk),d2.aneStage(w,c,Delta,AneBoundMlp,AneMlpK3i,AneAffineSplit,Plain,ANE_FRACTION,2048,shares,qk),a3.aneStage(w,c,Attention,AneBoundMlp,AneMlpK3i,AneAffineSplit,Plain,ANE_FRACTION,2048,shares),
|
|
542
|
+
d4.aneStage(w,c,Delta,AneBoundMlp,AneMlpK3i,AneAffineSplit,Plain,ANE_FRACTION,2048,shares,qk),d5.aneStage(w,c,Delta,AneBoundMlp,AneMlpK3i,AneAffineSplit,Plain,ANE_FRACTION,2048,shares,qk),d6.aneStage(w,c,Delta,AneBoundMlp,AneMlpK3i,AneAffineSplit,Plain,ANE_FRACTION,2048,shares,qk),a7.aneStage(w,c,Attention,AneBoundMlp,AneMlpK3i,AneAffineSplit,Plain,ANE_FRACTION,2048,shares),
|
|
543
|
+
d8.aneStage(w,c,Delta,AneBoundMlp,AneMlpK3i,AneAffineSplit,Plain,ANE_FRACTION,2048,shares,qk),d9.aneStage(w,c,Delta,AneBoundMlp,AneMlpK3i,AneAffineSplit,Plain,ANE_FRACTION,2048,shares,qk),d10.aneStage(w,c,Delta,AneBoundMlp,AneMlpK3i,AneAffineSplit,Plain,ANE_FRACTION,2048,shares,qk),a11.aneStage(w,c,Attention,AneBoundMlp,AneMlp,AneAffineSplit,Plain,ANE_FRACTION,2048,shares),
|
|
544
|
+
d12.aneStage(w,c,Delta,AneBoundMlp,AneMlpK3i,AneAffineSplit,Plain,ANE_FRACTION,2048,shares,qk),d13.aneStage(w,c,Delta,AneBoundMlp,AneMlp,AneAffineSplit,Plain,ANE_FRACTION,2048,shares,qk),d14.aneStage(w,c,Delta,AneBoundMlp,AneMlpK3i,AneAffineSplit,Plain,ANE_FRACTION,2048,shares,qk),a15.aneStage(w,c,Attention,AneBoundMlp,AneMlp,AneAffineSplit,Plain,ANE_FRACTION,2048,shares),
|
|
545
|
+
d16.aneStage(w,c,Delta,AneBoundMlp,AneMlp,AneAffineSplit,Plain,ANE_FRACTION,2048,shares,qk),d17.aneStage(w,c,Delta,AneBoundMlp,AneMlp,AneAffineSplit,Plain,ANE_FRACTION,2048,shares,qk),d18.aneStage(w,c,Delta,AneBoundMlp,AneMlp,AneAffineSplit,Plain,ANE_FRACTION,2048,shares,qk),a19.aneStage(w,c,Attention,AneBoundMlp,AneMlpK3i,AneAffineSplit,Plain,ANE_FRACTION,2048,shares),
|
|
546
|
+
d20.aneStage(w,c,Delta,AneBoundMlp,AneMlpK3i,AneAffineSplit,Plain,ANE_FRACTION,2048,shares,qk),d21.aneStage(w,c,Delta,AneBoundMlp,AneMlpK3i,AneAffineSplit,Plain,ANE_FRACTION,2048,shares,qk),d22.aneStage(w,c,Delta,AneBoundMlp,AneMlpK3i,AneAffineSplit,Plain,ANE_FRACTION,2048,shares,qk),a23.aneStage(w,c,Attention,AneBoundMlp,AneMlpK3i,AneAffineSplit,Plain,ANE_FRACTION,2048,shares),
|
|
547
|
+
d24.aneStage(w,c,Delta,AneBoundMlp,AneMlpK3i,AneAffineSplit,Plain,ANE_FRACTION,2048,shares,qk),d25.aneStage(w,c,Delta,AneBoundMlp,AneMlpK3i,AneAffineSplit,Plain,ANE_FRACTION,2048,shares,qk),d26.aneStage(w,c,Delta,AneBoundMlp,AneMlpK3i,AneAffineSplit,Plain,ANE_FRACTION,2048,shares,qk),a27.aneStage(w,c,Attention,AneBoundMlp,AneMlpK3i,AneAffineSplit,Plain,ANE_FRACTION,2048,shares),
|
|
548
|
+
d28.aneStage(w,c,Delta,AneBoundMlp,AneMlpK3i,AneAffineSplit,Plain,ANE_FRACTION,2048,shares,qk),d29.aneStage(w,c,Delta,AneBoundMlp,AneMlp,AneAffineSplit,Plain,ANE_FRACTION,2048,shares,qk),d30.aneStage(w,c,Delta,AneBoundMlp,AneMlpK3i,AneAffineSplit,Plain,ANE_FRACTION,2048,shares,qk),a31.aneStage(w,c,Attention,AneBoundMlp,AneMlp,AneAffineSplit,Plain,ANE_FRACTION,2048,shares),
|
|
549
|
+
d32.aneStage(w,c,Delta,AneBoundMlp,AneMlp,AneAffineSplit,Plain,ANE_FRACTION,2048,shares,qk),d33.aneStage(w,c,Delta,AneBoundMlp,AneMlp,AneAffineSplit,Plain,ANE_FRACTION,2048,shares,qk),d34.aneStage(w,c,Delta,AneBoundMlp,AneMlpK3i,AneAffineSplit,Plain,ANE_FRACTION,2048,shares,qk),a35.aneStage(w,c,Attention,AneBoundMlp,AneMlpK3i,AneAffineSplit,Plain,ANE_FRACTION,2048,shares),
|
|
550
|
+
d36.aneStage(w,c,Delta,AneBoundMlp,AneMlpK3i,AneAffineSplit,Plain,ANE_FRACTION,2048,shares,qk),d37.aneStage(w,c,Delta,AneBoundMlp,AneMlpK3i,AneAffineSplit,Plain,ANE_FRACTION,2048,shares,qk),d38.aneStage(w,c,Delta,AneBoundMlp,AneMlpK3i,AneAffineSplit,Plain,ANE_FRACTION,2048,shares,qk),a39.aneStage(w,c,Attention,AneBoundMlp,AneMlpK3i,AneAffineSplit,Plain,ANE_FRACTION,2048,shares),
|
|
551
|
+
d40.aneStage(w,c,Delta,AneBoundMlp,AneMlpK3i,AneAffineSplit,Plain,ANE_FRACTION,2048,shares,qk),d41.aneStage(w,c,Delta,AneBoundMlp,AneMlpK3i,AneAffineSplit,Plain,ANE_FRACTION,2048,shares,qk),d42.aneStage(w,c,Delta,AneBoundMlp,AneMlpK3i,AneAffineSplit,Plain,ANE_FRACTION,2048,shares,qk),a43.aneStage(w,c,Attention,AneBoundMlp,AneMlpK3i,AneAffineSplit,Plain,ANE_FRACTION,2048,shares),
|
|
552
|
+
d44.aneStage(w,c,Delta,AneBoundMlp,AneMlpK3i,AneAffineSplit,Plain,ANE_FRACTION,2048,shares,qk),d45.aneStage(w,c,Delta,AneBoundMlp,AneMlp,AneAffineSplit,Plain,ANE_FRACTION,2048,shares,qk),d46.aneStage(w,c,Delta,AneBoundMlp,AneMlp,AneAffineSplit,Plain,ANE_FRACTION,2048,shares,qk),a47.aneStage(w,c,Attention,AneBoundMlp,AneMlp,AneAffineSplit,Plain,ANE_FRACTION,2048,shares),
|
|
553
|
+
d48.aneStage(w,c,Delta,AneBoundMlp,AneMlp,AneAffineSplit,Plain,ANE_FRACTION,2048,shares,qk),d49.aneStage(w,c,Delta,AneBoundMlp,AneMlp,AneAffineSplit,Plain,ANE_FRACTION,2048,shares,qk),d50.aneStage(w,c,Delta,AneBoundMlp,AneMlpK3i,AneAffineSplit,Plain,ANE_FRACTION,2048,shares,qk),a51.aneStage(w,c,Attention,AneBoundMlp,AneMlpK3i,AneAffineSplit,Plain,ANE_FRACTION,2048,shares),
|
|
554
|
+
d52.aneStage(w,c,Delta,AneBoundMlp,AneMlpK3i,AneAffineSplit,Plain,ANE_FRACTION,2048,shares,qk),d53.aneStage(w,c,Delta,AneBoundMlp,AneMlpK3i,AneAffineSplit,Plain,ANE_FRACTION,2048,shares,qk),d54.aneStage(w,c,Delta,AneBoundMlp,AneMlp,AneAffineSplit,Plain,ANE_FRACTION,2048,shares,qk),a55.aneStage(w,c,Attention,AneBoundMlp,AneMlp,AneAffineSplit,Plain,ANE_FRACTION,2048,shares),
|
|
555
|
+
d56.aneStage(w,c,Delta,AneBoundMlp,AneMlp,AneAffineSplit,Plain,ANE_FRACTION,2048,shares,qk),d57.aneStage(w,c,Delta,AneBoundMlp,AneMlp,AneAffineSplit,Plain,ANE_FRACTION,2048,shares,qk),d58.aneStage(w,c,Delta,AneBoundMlp,AneMlp,AneAffineSplit,Plain,ANE_FRACTION,2048,shares,qk),a59.aneStage(w,c,Attention,AneBoundMlp,AneMlp,AneAffineSplit,Plain,ANE_FRACTION,2048,shares),
|
|
556
|
+
d60.aneStage(w,c,Delta,AneBoundMlp,AneMlp,AneAffineSplit,Plain,ANE_FRACTION,2048,shares,qk),d61.aneStage(w,c,Delta,AneBoundMlp,AneMlp,AneAffineSplit,Plain,ANE_FRACTION,2048,shares,qk),d62.aneStage(w,c,Delta,AneBoundMlp,AneMlpK3i,AneAffineSplit,Plain,ANE_FRACTION,2048,shares,qk),a63.aneStage(w,c,Attention,AneBoundMlp,AneMlpK3i,AneAffineSplit,Plain,ANE_FRACTION,2048,shares),
|
|
557
|
+
];
|
|
558
|
+
this.#ane1536=[
|
|
559
|
+
d0.aneStage(w,c,Delta,AneBoundMlp,AneMlp,AneAffineSplit,Plain,ANE_FRACTION,1536,shares,qk),d1.aneStage(w,c,Delta,AneBoundMlp,AneMlpK3i,AneAffineSplit,Plain,ANE_FRACTION,1536,shares,qk),d2.aneStage(w,c,Delta,AneBoundMlp,AneMlpK3i,AneAffineSplit,Plain,ANE_FRACTION,1536,shares,qk),a3.aneStage(w,c,Attention,AneBoundMlp,AneMlpK3i,AneAffineSplit,Plain,ANE_FRACTION,1536,shares),
|
|
560
|
+
d4.aneStage(w,c,Delta,AneBoundMlp,AneMlpK3i,AneAffineSplit,Plain,ANE_FRACTION,1536,shares,qk),d5.aneStage(w,c,Delta,AneBoundMlp,AneMlpK3i,AneAffineSplit,Plain,ANE_FRACTION,1536,shares,qk),d6.aneStage(w,c,Delta,AneBoundMlp,AneMlpK3i,AneAffineSplit,Plain,ANE_FRACTION,1536,shares,qk),a7.aneStage(w,c,Attention,AneBoundMlp,AneMlpK3i,AneAffineSplit,Plain,ANE_FRACTION,1536,shares),
|
|
561
|
+
d8.aneStage(w,c,Delta,AneBoundMlp,AneMlpK3i,AneAffineSplit,Plain,ANE_FRACTION,1536,shares,qk),d9.aneStage(w,c,Delta,AneBoundMlp,AneMlpK3i,AneAffineSplit,Plain,ANE_FRACTION,1536,shares,qk),d10.aneStage(w,c,Delta,AneBoundMlp,AneMlpK3i,AneAffineSplit,Plain,ANE_FRACTION,1536,shares,qk),a11.aneStage(w,c,Attention,AneBoundMlp,AneMlp,AneAffineSplit,Plain,ANE_FRACTION,1536,shares),
|
|
562
|
+
d12.aneStage(w,c,Delta,AneBoundMlp,AneMlpK3i,AneAffineSplit,Plain,ANE_FRACTION,1536,shares,qk),d13.aneStage(w,c,Delta,AneBoundMlp,AneMlp,AneAffineSplit,Plain,ANE_FRACTION,1536,shares,qk),d14.aneStage(w,c,Delta,AneBoundMlp,AneMlpK3i,AneAffineSplit,Plain,ANE_FRACTION,1536,shares,qk),a15.aneStage(w,c,Attention,AneBoundMlp,AneMlp,AneAffineSplit,Plain,ANE_FRACTION,1536,shares),
|
|
563
|
+
d16.aneStage(w,c,Delta,AneBoundMlp,AneMlp,AneAffineSplit,Plain,ANE_FRACTION,1536,shares,qk),d17.aneStage(w,c,Delta,AneBoundMlp,AneMlp,AneAffineSplit,Plain,ANE_FRACTION,1536,shares,qk),d18.aneStage(w,c,Delta,AneBoundMlp,AneMlp,AneAffineSplit,Plain,ANE_FRACTION,1536,shares,qk),a19.aneStage(w,c,Attention,AneBoundMlp,AneMlpK3i,AneAffineSplit,Plain,ANE_FRACTION,1536,shares),
|
|
564
|
+
d20.aneStage(w,c,Delta,AneBoundMlp,AneMlpK3i,AneAffineSplit,Plain,ANE_FRACTION,1536,shares,qk),d21.aneStage(w,c,Delta,AneBoundMlp,AneMlpK3i,AneAffineSplit,Plain,ANE_FRACTION,1536,shares,qk),d22.aneStage(w,c,Delta,AneBoundMlp,AneMlpK3i,AneAffineSplit,Plain,ANE_FRACTION,1536,shares,qk),a23.aneStage(w,c,Attention,AneBoundMlp,AneMlpK3i,AneAffineSplit,Plain,ANE_FRACTION,1536,shares),
|
|
565
|
+
d24.aneStage(w,c,Delta,AneBoundMlp,AneMlpK3i,AneAffineSplit,Plain,ANE_FRACTION,1536,shares,qk),d25.aneStage(w,c,Delta,AneBoundMlp,AneMlpK3i,AneAffineSplit,Plain,ANE_FRACTION,1536,shares,qk),d26.aneStage(w,c,Delta,AneBoundMlp,AneMlpK3i,AneAffineSplit,Plain,ANE_FRACTION,1536,shares,qk),a27.aneStage(w,c,Attention,AneBoundMlp,AneMlpK3i,AneAffineSplit,Plain,ANE_FRACTION,1536,shares),
|
|
566
|
+
d28.aneStage(w,c,Delta,AneBoundMlp,AneMlpK3i,AneAffineSplit,Plain,ANE_FRACTION,1536,shares,qk),d29.aneStage(w,c,Delta,AneBoundMlp,AneMlp,AneAffineSplit,Plain,ANE_FRACTION,1536,shares,qk),d30.aneStage(w,c,Delta,AneBoundMlp,AneMlpK3i,AneAffineSplit,Plain,ANE_FRACTION,1536,shares,qk),a31.aneStage(w,c,Attention,AneBoundMlp,AneMlp,AneAffineSplit,Plain,ANE_FRACTION,1536,shares),
|
|
567
|
+
d32.aneStage(w,c,Delta,AneBoundMlp,AneMlp,AneAffineSplit,Plain,ANE_FRACTION,1536,shares,qk),d33.aneStage(w,c,Delta,AneBoundMlp,AneMlp,AneAffineSplit,Plain,ANE_FRACTION,1536,shares,qk),d34.aneStage(w,c,Delta,AneBoundMlp,AneMlpK3i,AneAffineSplit,Plain,ANE_FRACTION,1536,shares,qk),a35.aneStage(w,c,Attention,AneBoundMlp,AneMlpK3i,AneAffineSplit,Plain,ANE_FRACTION,1536,shares),
|
|
568
|
+
d36.aneStage(w,c,Delta,AneBoundMlp,AneMlpK3i,AneAffineSplit,Plain,ANE_FRACTION,1536,shares,qk),d37.aneStage(w,c,Delta,AneBoundMlp,AneMlpK3i,AneAffineSplit,Plain,ANE_FRACTION,1536,shares,qk),d38.aneStage(w,c,Delta,AneBoundMlp,AneMlpK3i,AneAffineSplit,Plain,ANE_FRACTION,1536,shares,qk),a39.aneStage(w,c,Attention,AneBoundMlp,AneMlpK3i,AneAffineSplit,Plain,ANE_FRACTION,1536,shares),
|
|
569
|
+
d40.aneStage(w,c,Delta,AneBoundMlp,AneMlpK3i,AneAffineSplit,Plain,ANE_FRACTION,1536,shares,qk),d41.aneStage(w,c,Delta,AneBoundMlp,AneMlpK3i,AneAffineSplit,Plain,ANE_FRACTION,1536,shares,qk),d42.aneStage(w,c,Delta,AneBoundMlp,AneMlpK3i,AneAffineSplit,Plain,ANE_FRACTION,1536,shares,qk),a43.aneStage(w,c,Attention,AneBoundMlp,AneMlpK3i,AneAffineSplit,Plain,ANE_FRACTION,1536,shares),
|
|
570
|
+
d44.aneStage(w,c,Delta,AneBoundMlp,AneMlpK3i,AneAffineSplit,Plain,ANE_FRACTION,1536,shares,qk),d45.aneStage(w,c,Delta,AneBoundMlp,AneMlp,AneAffineSplit,Plain,ANE_FRACTION,1536,shares,qk),d46.aneStage(w,c,Delta,AneBoundMlp,AneMlp,AneAffineSplit,Plain,ANE_FRACTION,1536,shares,qk),a47.aneStage(w,c,Attention,AneBoundMlp,AneMlp,AneAffineSplit,Plain,ANE_FRACTION,1536,shares),
|
|
571
|
+
d48.aneStage(w,c,Delta,AneBoundMlp,AneMlp,AneAffineSplit,Plain,ANE_FRACTION,1536,shares,qk),d49.aneStage(w,c,Delta,AneBoundMlp,AneMlp,AneAffineSplit,Plain,ANE_FRACTION,1536,shares,qk),d50.aneStage(w,c,Delta,AneBoundMlp,AneMlpK3i,AneAffineSplit,Plain,ANE_FRACTION,1536,shares,qk),a51.aneStage(w,c,Attention,AneBoundMlp,AneMlpK3i,AneAffineSplit,Plain,ANE_FRACTION,1536,shares),
|
|
572
|
+
d52.aneStage(w,c,Delta,AneBoundMlp,AneMlpK3i,AneAffineSplit,Plain,ANE_FRACTION,1536,shares,qk),d53.aneStage(w,c,Delta,AneBoundMlp,AneMlpK3i,AneAffineSplit,Plain,ANE_FRACTION,1536,shares,qk),d54.aneStage(w,c,Delta,AneBoundMlp,AneMlp,AneAffineSplit,Plain,ANE_FRACTION,1536,shares,qk),a55.aneStage(w,c,Attention,AneBoundMlp,AneMlp,AneAffineSplit,Plain,ANE_FRACTION,1536,shares),
|
|
573
|
+
d56.aneStage(w,c,Delta,AneBoundMlp,AneMlp,AneAffineSplit,Plain,ANE_FRACTION,1536,shares,qk),d57.aneStage(w,c,Delta,AneBoundMlp,AneMlp,AneAffineSplit,Plain,ANE_FRACTION,1536,shares,qk),d58.aneStage(w,c,Delta,AneBoundMlp,AneMlp,AneAffineSplit,Plain,ANE_FRACTION,1536,shares,qk),a59.aneStage(w,c,Attention,AneBoundMlp,AneMlp,AneAffineSplit,Plain,ANE_FRACTION,1536,shares),
|
|
574
|
+
d60.aneStage(w,c,Delta,AneBoundMlp,AneMlp,AneAffineSplit,Plain,ANE_FRACTION,1536,shares,qk),d61.aneStage(w,c,Delta,AneBoundMlp,AneMlp,AneAffineSplit,Plain,ANE_FRACTION,1536,shares,qk),d62.aneStage(w,c,Delta,AneBoundMlp,AneMlpK3i,AneAffineSplit,Plain,ANE_FRACTION,1536,shares,qk),a63.aneStage(w,c,Attention,AneBoundMlp,AneMlpK3i,AneAffineSplit,Plain,ANE_FRACTION,1536,shares),
|
|
575
|
+
];
|
|
576
|
+
this.#ane1024=[
|
|
577
|
+
d0.aneStage(w,c,Delta,AneBoundMlp,AneMlp,AneAffineSplit,Plain,ANE_FRACTION,1024,shares,qk),d1.aneStage(w,c,Delta,AneBoundMlp,AneMlpK3i,AneAffineSplit,Plain,ANE_FRACTION,1024,shares,qk),d2.aneStage(w,c,Delta,AneBoundMlp,AneMlpK3i,AneAffineSplit,Plain,ANE_FRACTION,1024,shares,qk),a3.aneStage(w,c,Attention,AneBoundMlp,AneMlpK3i,AneAffineSplit,Plain,ANE_FRACTION,1024,shares),
|
|
578
|
+
d4.aneStage(w,c,Delta,AneBoundMlp,AneMlpK3i,AneAffineSplit,Plain,ANE_FRACTION,1024,shares,qk),d5.aneStage(w,c,Delta,AneBoundMlp,AneMlpK3i,AneAffineSplit,Plain,ANE_FRACTION,1024,shares,qk),d6.aneStage(w,c,Delta,AneBoundMlp,AneMlpK3i,AneAffineSplit,Plain,ANE_FRACTION,1024,shares,qk),a7.aneStage(w,c,Attention,AneBoundMlp,AneMlpK3i,AneAffineSplit,Plain,ANE_FRACTION,1024,shares),
|
|
579
|
+
d8.aneStage(w,c,Delta,AneBoundMlp,AneMlpK3i,AneAffineSplit,Plain,ANE_FRACTION,1024,shares,qk),d9.aneStage(w,c,Delta,AneBoundMlp,AneMlpK3i,AneAffineSplit,Plain,ANE_FRACTION,1024,shares,qk),d10.aneStage(w,c,Delta,AneBoundMlp,AneMlpK3i,AneAffineSplit,Plain,ANE_FRACTION,1024,shares,qk),a11.aneStage(w,c,Attention,AneBoundMlp,AneMlp,AneAffineSplit,Plain,ANE_FRACTION,1024,shares),
|
|
580
|
+
d12.aneStage(w,c,Delta,AneBoundMlp,AneMlpK3i,AneAffineSplit,Plain,ANE_FRACTION,1024,shares,qk),d13.aneStage(w,c,Delta,AneBoundMlp,AneMlp,AneAffineSplit,Plain,ANE_FRACTION,1024,shares,qk),d14.aneStage(w,c,Delta,AneBoundMlp,AneMlpK3i,AneAffineSplit,Plain,ANE_FRACTION,1024,shares,qk),a15.aneStage(w,c,Attention,AneBoundMlp,AneMlp,AneAffineSplit,Plain,ANE_FRACTION,1024,shares),
|
|
581
|
+
d16.aneStage(w,c,Delta,AneBoundMlp,AneMlp,AneAffineSplit,Plain,ANE_FRACTION,1024,shares,qk),d17.aneStage(w,c,Delta,AneBoundMlp,AneMlp,AneAffineSplit,Plain,ANE_FRACTION,1024,shares,qk),d18.aneStage(w,c,Delta,AneBoundMlp,AneMlp,AneAffineSplit,Plain,ANE_FRACTION,1024,shares,qk),a19.aneStage(w,c,Attention,AneBoundMlp,AneMlpK3i,AneAffineSplit,Plain,ANE_FRACTION,1024,shares),
|
|
582
|
+
d20.aneStage(w,c,Delta,AneBoundMlp,AneMlpK3i,AneAffineSplit,Plain,ANE_FRACTION,1024,shares,qk),d21.aneStage(w,c,Delta,AneBoundMlp,AneMlpK3i,AneAffineSplit,Plain,ANE_FRACTION,1024,shares,qk),d22.aneStage(w,c,Delta,AneBoundMlp,AneMlpK3i,AneAffineSplit,Plain,ANE_FRACTION,1024,shares,qk),a23.aneStage(w,c,Attention,AneBoundMlp,AneMlpK3i,AneAffineSplit,Plain,ANE_FRACTION,1024,shares),
|
|
583
|
+
d24.aneStage(w,c,Delta,AneBoundMlp,AneMlpK3i,AneAffineSplit,Plain,ANE_FRACTION,1024,shares,qk),d25.aneStage(w,c,Delta,AneBoundMlp,AneMlpK3i,AneAffineSplit,Plain,ANE_FRACTION,1024,shares,qk),d26.aneStage(w,c,Delta,AneBoundMlp,AneMlpK3i,AneAffineSplit,Plain,ANE_FRACTION,1024,shares,qk),a27.aneStage(w,c,Attention,AneBoundMlp,AneMlpK3i,AneAffineSplit,Plain,ANE_FRACTION,1024,shares),
|
|
584
|
+
d28.aneStage(w,c,Delta,AneBoundMlp,AneMlpK3i,AneAffineSplit,Plain,ANE_FRACTION,1024,shares,qk),d29.aneStage(w,c,Delta,AneBoundMlp,AneMlp,AneAffineSplit,Plain,ANE_FRACTION,1024,shares,qk),d30.aneStage(w,c,Delta,AneBoundMlp,AneMlpK3i,AneAffineSplit,Plain,ANE_FRACTION,1024,shares,qk),a31.aneStage(w,c,Attention,AneBoundMlp,AneMlp,AneAffineSplit,Plain,ANE_FRACTION,1024,shares),
|
|
585
|
+
d32.aneStage(w,c,Delta,AneBoundMlp,AneMlp,AneAffineSplit,Plain,ANE_FRACTION,1024,shares,qk),d33.aneStage(w,c,Delta,AneBoundMlp,AneMlp,AneAffineSplit,Plain,ANE_FRACTION,1024,shares,qk),d34.aneStage(w,c,Delta,AneBoundMlp,AneMlpK3i,AneAffineSplit,Plain,ANE_FRACTION,1024,shares,qk),a35.aneStage(w,c,Attention,AneBoundMlp,AneMlpK3i,AneAffineSplit,Plain,ANE_FRACTION,1024,shares),
|
|
586
|
+
d36.aneStage(w,c,Delta,AneBoundMlp,AneMlpK3i,AneAffineSplit,Plain,ANE_FRACTION,1024,shares,qk),d37.aneStage(w,c,Delta,AneBoundMlp,AneMlpK3i,AneAffineSplit,Plain,ANE_FRACTION,1024,shares,qk),d38.aneStage(w,c,Delta,AneBoundMlp,AneMlpK3i,AneAffineSplit,Plain,ANE_FRACTION,1024,shares,qk),a39.aneStage(w,c,Attention,AneBoundMlp,AneMlpK3i,AneAffineSplit,Plain,ANE_FRACTION,1024,shares),
|
|
587
|
+
d40.aneStage(w,c,Delta,AneBoundMlp,AneMlpK3i,AneAffineSplit,Plain,ANE_FRACTION,1024,shares,qk),d41.aneStage(w,c,Delta,AneBoundMlp,AneMlpK3i,AneAffineSplit,Plain,ANE_FRACTION,1024,shares,qk),d42.aneStage(w,c,Delta,AneBoundMlp,AneMlpK3i,AneAffineSplit,Plain,ANE_FRACTION,1024,shares,qk),a43.aneStage(w,c,Attention,AneBoundMlp,AneMlpK3i,AneAffineSplit,Plain,ANE_FRACTION,1024,shares),
|
|
588
|
+
d44.aneStage(w,c,Delta,AneBoundMlp,AneMlpK3i,AneAffineSplit,Plain,ANE_FRACTION,1024,shares,qk),d45.aneStage(w,c,Delta,AneBoundMlp,AneMlp,AneAffineSplit,Plain,ANE_FRACTION,1024,shares,qk),d46.aneStage(w,c,Delta,AneBoundMlp,AneMlp,AneAffineSplit,Plain,ANE_FRACTION,1024,shares,qk),a47.aneStage(w,c,Attention,AneBoundMlp,AneMlp,AneAffineSplit,Plain,ANE_FRACTION,1024,shares),
|
|
589
|
+
d48.aneStage(w,c,Delta,AneBoundMlp,AneMlp,AneAffineSplit,Plain,ANE_FRACTION,1024,shares,qk),d49.aneStage(w,c,Delta,AneBoundMlp,AneMlp,AneAffineSplit,Plain,ANE_FRACTION,1024,shares,qk),d50.aneStage(w,c,Delta,AneBoundMlp,AneMlpK3i,AneAffineSplit,Plain,ANE_FRACTION,1024,shares,qk),a51.aneStage(w,c,Attention,AneBoundMlp,AneMlpK3i,AneAffineSplit,Plain,ANE_FRACTION,1024,shares),
|
|
590
|
+
d52.aneStage(w,c,Delta,AneBoundMlp,AneMlpK3i,AneAffineSplit,Plain,ANE_FRACTION,1024,shares,qk),d53.aneStage(w,c,Delta,AneBoundMlp,AneMlpK3i,AneAffineSplit,Plain,ANE_FRACTION,1024,shares,qk),d54.aneStage(w,c,Delta,AneBoundMlp,AneMlp,AneAffineSplit,Plain,ANE_FRACTION,1024,shares,qk),a55.aneStage(w,c,Attention,AneBoundMlp,AneMlp,AneAffineSplit,Plain,ANE_FRACTION,1024,shares),
|
|
591
|
+
d56.aneStage(w,c,Delta,AneBoundMlp,AneMlp,AneAffineSplit,Plain,ANE_FRACTION,1024,shares,qk),d57.aneStage(w,c,Delta,AneBoundMlp,AneMlp,AneAffineSplit,Plain,ANE_FRACTION,1024,shares,qk),d58.aneStage(w,c,Delta,AneBoundMlp,AneMlp,AneAffineSplit,Plain,ANE_FRACTION,1024,shares,qk),a59.aneStage(w,c,Attention,AneBoundMlp,AneMlp,AneAffineSplit,Plain,ANE_FRACTION,1024,shares),
|
|
592
|
+
d60.aneStage(w,c,Delta,AneBoundMlp,AneMlp,AneAffineSplit,Plain,ANE_FRACTION,1024,shares,qk),d61.aneStage(w,c,Delta,AneBoundMlp,AneMlp,AneAffineSplit,Plain,ANE_FRACTION,1024,shares,qk),d62.aneStage(w,c,Delta,AneBoundMlp,AneMlpK3i,AneAffineSplit,Plain,ANE_FRACTION,1024,shares,qk),a63.aneStage(w,c,Attention,AneBoundMlp,AneMlpK3i,AneAffineSplit,Plain,ANE_FRACTION,1024,shares),
|
|
593
|
+
];
|
|
594
|
+
// GPU chunks, media and known appends retain the stock Trellis arithmetic.
|
|
595
|
+
this.#stockNarrow=[
|
|
596
|
+
d0.fused(w,c,Delta,TrellisFusedMLP,Scatter,Plain,qk),d1.fused(w,c,Delta,TrellisFusedMLP,Scatter,Plain,qk),d2.fused(w,c,Delta,TrellisFusedMLP,Scatter,Plain,qk),a3.fused(w,c,Attention,TrellisFusedMLP,Scatter,Plain),
|
|
597
|
+
d4.fused(w,c,Delta,TrellisFusedMLP,Scatter,Plain,qk),d5.fused(w,c,Delta,TrellisFusedMLP,Scatter,Plain,qk),d6.fused(w,c,Delta,TrellisFusedMLP,Scatter,Plain,qk),a7.fused(w,c,Attention,TrellisFusedMLP,Scatter,Plain),
|
|
598
|
+
d8.fused(w,c,Delta,TrellisFusedMLP,Scatter,Plain,qk),d9.fused(w,c,Delta,TrellisFusedMLP,Scatter,Plain,qk),d10.fused(w,c,Delta,TrellisFusedMLP,Scatter,Plain,qk),a11.fused(w,c,Attention,TrellisFusedMLP,Scatter,Plain),
|
|
599
|
+
d12.fused(w,c,Delta,TrellisFusedMLP,Scatter,Plain,qk),d13.fused(w,c,Delta,TrellisFusedMLP,Scatter,Plain,qk),d14.fused(w,c,Delta,TrellisFusedMLP,Scatter,Plain,qk),a15.prompt(w,c,Attention,TrellisUnfusedMLP,Reduce,Scatter,Plain),
|
|
600
|
+
d16.fused(w,c,Delta,TrellisFusedMLP,Scatter,Plain,qk),d17.fused(w,c,Delta,TrellisFusedMLP,Scatter,Plain,qk),d18.fused(w,c,Delta,TrellisFusedMLP,Scatter,Plain,qk),a19.fused(w,c,Attention,TrellisFusedMLP,Scatter,Plain),
|
|
601
|
+
d20.fused(w,c,Delta,TrellisFusedMLP,Scatter,Plain,qk),d21.prompt(w,c,Delta,TrellisUnfusedMLP,Reduce,Scatter,Plain,qk),d22.fused(w,c,Delta,TrellisFusedMLP,Scatter,Plain,qk),a23.fused(w,c,Attention,TrellisFusedMLP,Scatter,Plain),
|
|
602
|
+
d24.fused(w,c,Delta,TrellisFusedMLP,Scatter,Plain,qk),d25.fused(w,c,Delta,TrellisFusedMLP,Scatter,Plain,qk),d26.fused(w,c,Delta,TrellisFusedMLP,Scatter,Plain,qk),a27.fused(w,c,Attention,TrellisFusedMLP,Scatter,Plain),
|
|
603
|
+
d28.fused(w,c,Delta,TrellisFusedMLP,Scatter,Plain,qk),d29.fused(w,c,Delta,TrellisFusedMLP,Scatter,Plain,qk),d30.fused(w,c,Delta,TrellisFusedMLP,Scatter,Plain,qk),a31.fused(w,c,Attention,TrellisFusedMLP,Scatter,Plain),
|
|
604
|
+
d32.fused(w,c,Delta,TrellisFusedMLP,Scatter,Plain,qk),d33.fused(w,c,Delta,TrellisFusedMLP,Scatter,Plain,qk),d34.fused(w,c,Delta,TrellisFusedMLP,Scatter,Plain,qk),a35.prompt(w,c,Attention,TrellisUnfusedMLP,Reduce,Scatter,Plain),
|
|
605
|
+
d36.prompt(w,c,Delta,TrellisUnfusedMLP,Reduce,Scatter,Plain,qk),d37.fused(w,c,Delta,TrellisFusedMLP,Scatter,Plain,qk),d38.fused(w,c,Delta,TrellisFusedMLP,Scatter,Plain,qk),a39.fused(w,c,Attention,TrellisFusedMLP,Scatter,Plain),
|
|
606
|
+
d40.fused(w,c,Delta,TrellisFusedMLP,Scatter,Plain,qk),d41.fused(w,c,Delta,TrellisFusedMLP,Scatter,Plain,qk),d42.fused(w,c,Delta,TrellisFusedMLP,Scatter,Plain,qk),a43.prompt(w,c,Attention,TrellisUnfusedMLP,Reduce,Scatter,Plain),
|
|
607
|
+
d44.fused(w,c,Delta,TrellisFusedMLP,Scatter,Plain,qk),d45.fused(w,c,Delta,TrellisFusedMLP,Scatter,Plain,qk),d46.fused(w,c,Delta,TrellisFusedMLP,Scatter,Plain,qk),a47.fused(w,c,Attention,TrellisFusedMLP,Scatter,Plain),
|
|
608
|
+
d48.fused(w,c,Delta,TrellisFusedMLP,Scatter,Plain,qk),d49.fused(w,c,Delta,TrellisFusedMLP,Scatter,Plain,qk),d50.prompt(w,c,Delta,TrellisUnfusedMLP,Reduce,Scatter,Plain,qk),a51.prompt(w,c,Attention,TrellisUnfusedMLP,Reduce,Scatter,Plain),
|
|
609
|
+
d52.fused(w,c,Delta,TrellisFusedMLP,Scatter,Plain,qk),d53.prompt(w,c,Delta,TrellisUnfusedMLP,Reduce,Scatter,Plain,qk),d54.fused(w,c,Delta,TrellisFusedMLP,Scatter,Plain,qk),a55.fused(w,c,Attention,TrellisFusedMLP,Scatter,Plain),
|
|
610
|
+
d56.prompt(w,c,Delta,TrellisUnfusedMLP,Reduce,Scatter,Plain,qk),d57.fused(w,c,Delta,TrellisFusedMLP,Scatter,Plain,qk),d58.fused(w,c,Delta,TrellisFusedMLP,Scatter,Plain,qk),a59.fused(w,c,Attention,TrellisFusedMLP,Scatter,Plain),
|
|
611
|
+
d60.fused(w,c,Delta,TrellisFusedMLP,Scatter,Plain,qk),d61.fused(w,c,Delta,TrellisFusedMLP,Scatter,Plain,qk),d62.prompt(w,c,Delta,TrellisUnfusedMLP,Reduce,Scatter,Plain,qk),a63.fused(w,c,Attention,TrellisFusedMLP,Scatter,Plain),
|
|
612
|
+
];
|
|
613
|
+
this.#stockFiveToEight=[
|
|
614
|
+
d0.prompt(w,c,Delta,TrellisUnfusedMLP,Wide,SplitK,Plain,qk),d1.prompt(w,c,Delta,TrellisUnfusedMLP,Wide,SplitK,Plain,qk),d2.prompt(w,c,Delta,TrellisUnfusedMLP,Wide,SplitK,Plain,qk),a3.prompt(w,c,Attention,TrellisUnfusedMLP,Wide,SplitK,Plain),
|
|
615
|
+
d4.prompt(w,c,Delta,TrellisUnfusedMLP,Wide,SplitK,Plain,qk),d5.prompt(w,c,Delta,TrellisUnfusedMLP,Wide,SplitK,Plain,qk),d6.prompt(w,c,Delta,TrellisUnfusedMLP,Wide,SplitK,Plain,qk),a7.prompt(w,c,Attention,TrellisUnfusedMLP,Wide,SplitK,Plain),
|
|
616
|
+
d8.prompt(w,c,Delta,TrellisUnfusedMLP,Wide,SplitK,Plain,qk),d9.prompt(w,c,Delta,TrellisUnfusedMLP,Wide,SplitK,Plain,qk),d10.prompt(w,c,Delta,TrellisUnfusedMLP,Wide,SplitK,Plain,qk),a11.prompt(w,c,Attention,TrellisUnfusedMLP,Wide,SplitK,Plain),
|
|
617
|
+
d12.prompt(w,c,Delta,TrellisUnfusedMLP,Wide,SplitK,Plain,qk),d13.prompt(w,c,Delta,TrellisUnfusedMLP,Wide,SplitK,Plain,qk),d14.prompt(w,c,Delta,TrellisUnfusedMLP,Wide,SplitK,Plain,qk),a15.prompt(w,c,Attention,TrellisUnfusedMLP,Wide,SplitK,Plain),
|
|
618
|
+
d16.prompt(w,c,Delta,TrellisUnfusedMLP,Wide,SplitK,Plain,qk),d17.prompt(w,c,Delta,TrellisUnfusedMLP,Wide,SplitK,Plain,qk),d18.prompt(w,c,Delta,TrellisUnfusedMLP,Wide,SplitK,Plain,qk),a19.prompt(w,c,Attention,TrellisUnfusedMLP,Wide,SplitK,Plain),
|
|
619
|
+
d20.prompt(w,c,Delta,TrellisUnfusedMLP,Wide,SplitK,Plain,qk),d21.prompt(w,c,Delta,TrellisUnfusedMLP,Wide,SplitK,Plain,qk),d22.prompt(w,c,Delta,TrellisUnfusedMLP,Wide,SplitK,Plain,qk),a23.prompt(w,c,Attention,TrellisUnfusedMLP,Wide,SplitK,Plain),
|
|
620
|
+
d24.prompt(w,c,Delta,TrellisUnfusedMLP,Wide,SplitK,Plain,qk),d25.prompt(w,c,Delta,TrellisUnfusedMLP,Wide,SplitK,Plain,qk),d26.prompt(w,c,Delta,TrellisUnfusedMLP,Wide,SplitK,Plain,qk),a27.prompt(w,c,Attention,TrellisUnfusedMLP,Wide,SplitK,Plain),
|
|
621
|
+
d28.prompt(w,c,Delta,TrellisUnfusedMLP,Wide,SplitK,Plain,qk),d29.prompt(w,c,Delta,TrellisUnfusedMLP,Wide,SplitK,Plain,qk),d30.prompt(w,c,Delta,TrellisUnfusedMLP,Wide,SplitK,Plain,qk),a31.prompt(w,c,Attention,TrellisUnfusedMLP,Wide,SplitK,Plain),
|
|
622
|
+
d32.prompt(w,c,Delta,TrellisUnfusedMLP,Wide,SplitK,Plain,qk),d33.prompt(w,c,Delta,TrellisUnfusedMLP,Wide,SplitK,Plain,qk),d34.prompt(w,c,Delta,TrellisUnfusedMLP,Wide,SplitK,Plain,qk),a35.prompt(w,c,Attention,TrellisUnfusedMLP,Wide,SplitK,Plain),
|
|
623
|
+
d36.prompt(w,c,Delta,TrellisUnfusedMLP,Wide,SplitK,Plain,qk),d37.prompt(w,c,Delta,TrellisUnfusedMLP,Wide,SplitK,Plain,qk),d38.prompt(w,c,Delta,TrellisUnfusedMLP,Wide,SplitK,Plain,qk),a39.prompt(w,c,Attention,TrellisUnfusedMLP,Wide,SplitK,Plain),
|
|
624
|
+
d40.prompt(w,c,Delta,TrellisUnfusedMLP,Wide,SplitK,Plain,qk),d41.prompt(w,c,Delta,TrellisUnfusedMLP,Wide,SplitK,Plain,qk),d42.prompt(w,c,Delta,TrellisUnfusedMLP,Wide,SplitK,Plain,qk),a43.prompt(w,c,Attention,TrellisUnfusedMLP,Wide,SplitK,Plain),
|
|
625
|
+
d44.prompt(w,c,Delta,TrellisUnfusedMLP,Wide,SplitK,Plain,qk),d45.prompt(w,c,Delta,TrellisUnfusedMLP,Wide,SplitK,Plain,qk),d46.prompt(w,c,Delta,TrellisUnfusedMLP,Wide,SplitK,Plain,qk),a47.prompt(w,c,Attention,TrellisUnfusedMLP,Wide,SplitK,Plain),
|
|
626
|
+
d48.prompt(w,c,Delta,TrellisUnfusedMLP,Wide,SplitK,Plain,qk),d49.prompt(w,c,Delta,TrellisUnfusedMLP,Wide,SplitK,Plain,qk),d50.prompt(w,c,Delta,TrellisUnfusedMLP,Wide,SplitK,Plain,qk),a51.prompt(w,c,Attention,TrellisUnfusedMLP,Wide,SplitK,Plain),
|
|
627
|
+
d52.prompt(w,c,Delta,TrellisUnfusedMLP,Wide,SplitK,Plain,qk),d53.prompt(w,c,Delta,TrellisUnfusedMLP,Wide,SplitK,Plain,qk),d54.prompt(w,c,Delta,TrellisUnfusedMLP,Wide,SplitK,Plain,qk),a55.prompt(w,c,Attention,TrellisUnfusedMLP,Wide,SplitK,Plain),
|
|
628
|
+
d56.prompt(w,c,Delta,TrellisUnfusedMLP,Wide,SplitK,Plain,qk),d57.prompt(w,c,Delta,TrellisUnfusedMLP,Wide,SplitK,Plain,qk),d58.prompt(w,c,Delta,TrellisUnfusedMLP,Wide,SplitK,Plain,qk),a59.prompt(w,c,Attention,TrellisUnfusedMLP,Wide,SplitK,Plain),
|
|
629
|
+
d60.prompt(w,c,Delta,TrellisUnfusedMLP,Wide,SplitK,Plain,qk),d61.prompt(w,c,Delta,TrellisUnfusedMLP,Wide,SplitK,Plain,qk),d62.prompt(w,c,Delta,TrellisUnfusedMLP,Wide,SplitK,Plain,qk),a63.prompt(w,c,Attention,TrellisUnfusedMLP,Wide,SplitK,Plain),
|
|
630
|
+
];
|
|
631
|
+
this.#knownNarrow=[
|
|
632
|
+
d0.withMlp(w,c,Delta,this.#stockNarrow[0].mlp,SeparateRows,qk),d1.withMlp(w,c,Delta,this.#stockNarrow[1].mlp,SeparateRows,qk),d2.withMlp(w,c,Delta,this.#stockNarrow[2].mlp,SeparateRows,qk),a3.withMlp(w,c,Attention,this.#stockNarrow[3].mlp,SeparateRows),
|
|
633
|
+
d4.withMlp(w,c,Delta,this.#stockNarrow[4].mlp,SeparateRows,qk),d5.withMlp(w,c,Delta,this.#stockNarrow[5].mlp,SeparateRows,qk),d6.withMlp(w,c,Delta,this.#stockNarrow[6].mlp,SeparateRows,qk),a7.withMlp(w,c,Attention,this.#stockNarrow[7].mlp,SeparateRows),
|
|
634
|
+
d8.withMlp(w,c,Delta,this.#stockNarrow[8].mlp,SeparateRows,qk),d9.withMlp(w,c,Delta,this.#stockNarrow[9].mlp,SeparateRows,qk),d10.withMlp(w,c,Delta,this.#stockNarrow[10].mlp,SeparateRows,qk),a11.withMlp(w,c,Attention,this.#stockNarrow[11].mlp,SeparateRows),
|
|
635
|
+
d12.withMlp(w,c,Delta,this.#stockNarrow[12].mlp,SeparateRows,qk),d13.withMlp(w,c,Delta,this.#stockNarrow[13].mlp,SeparateRows,qk),d14.withMlp(w,c,Delta,this.#stockNarrow[14].mlp,SeparateRows,qk),a15.withMlp(w,c,Attention,this.#stockNarrow[15].mlp,SeparateRows),
|
|
636
|
+
d16.withMlp(w,c,Delta,this.#stockNarrow[16].mlp,SeparateRows,qk),d17.withMlp(w,c,Delta,this.#stockNarrow[17].mlp,SeparateRows,qk),d18.withMlp(w,c,Delta,this.#stockNarrow[18].mlp,SeparateRows,qk),a19.withMlp(w,c,Attention,this.#stockNarrow[19].mlp,SeparateRows),
|
|
637
|
+
d20.withMlp(w,c,Delta,this.#stockNarrow[20].mlp,SeparateRows,qk),d21.withMlp(w,c,Delta,this.#stockNarrow[21].mlp,SeparateRows,qk),d22.withMlp(w,c,Delta,this.#stockNarrow[22].mlp,SeparateRows,qk),a23.withMlp(w,c,Attention,this.#stockNarrow[23].mlp,SeparateRows),
|
|
638
|
+
d24.withMlp(w,c,Delta,this.#stockNarrow[24].mlp,SeparateRows,qk),d25.withMlp(w,c,Delta,this.#stockNarrow[25].mlp,SeparateRows,qk),d26.withMlp(w,c,Delta,this.#stockNarrow[26].mlp,SeparateRows,qk),a27.withMlp(w,c,Attention,this.#stockNarrow[27].mlp,SeparateRows),
|
|
639
|
+
d28.withMlp(w,c,Delta,this.#stockNarrow[28].mlp,SeparateRows,qk),d29.withMlp(w,c,Delta,this.#stockNarrow[29].mlp,SeparateRows,qk),d30.withMlp(w,c,Delta,this.#stockNarrow[30].mlp,SeparateRows,qk),a31.withMlp(w,c,Attention,this.#stockNarrow[31].mlp,SeparateRows),
|
|
640
|
+
d32.withMlp(w,c,Delta,this.#stockNarrow[32].mlp,SeparateRows,qk),d33.withMlp(w,c,Delta,this.#stockNarrow[33].mlp,SeparateRows,qk),d34.withMlp(w,c,Delta,this.#stockNarrow[34].mlp,SeparateRows,qk),a35.withMlp(w,c,Attention,this.#stockNarrow[35].mlp,SeparateRows),
|
|
641
|
+
d36.withMlp(w,c,Delta,this.#stockNarrow[36].mlp,SeparateRows,qk),d37.withMlp(w,c,Delta,this.#stockNarrow[37].mlp,SeparateRows,qk),d38.withMlp(w,c,Delta,this.#stockNarrow[38].mlp,SeparateRows,qk),a39.withMlp(w,c,Attention,this.#stockNarrow[39].mlp,SeparateRows),
|
|
642
|
+
d40.withMlp(w,c,Delta,this.#stockNarrow[40].mlp,SeparateRows,qk),d41.withMlp(w,c,Delta,this.#stockNarrow[41].mlp,SeparateRows,qk),d42.withMlp(w,c,Delta,this.#stockNarrow[42].mlp,SeparateRows,qk),a43.withMlp(w,c,Attention,this.#stockNarrow[43].mlp,SeparateRows),
|
|
643
|
+
d44.withMlp(w,c,Delta,this.#stockNarrow[44].mlp,SeparateRows,qk),d45.withMlp(w,c,Delta,this.#stockNarrow[45].mlp,SeparateRows,qk),d46.withMlp(w,c,Delta,this.#stockNarrow[46].mlp,SeparateRows,qk),a47.withMlp(w,c,Attention,this.#stockNarrow[47].mlp,SeparateRows),
|
|
644
|
+
d48.withMlp(w,c,Delta,this.#stockNarrow[48].mlp,SeparateRows,qk),d49.withMlp(w,c,Delta,this.#stockNarrow[49].mlp,SeparateRows,qk),d50.withMlp(w,c,Delta,this.#stockNarrow[50].mlp,SeparateRows,qk),a51.withMlp(w,c,Attention,this.#stockNarrow[51].mlp,SeparateRows),
|
|
645
|
+
d52.withMlp(w,c,Delta,this.#stockNarrow[52].mlp,SeparateRows,qk),d53.withMlp(w,c,Delta,this.#stockNarrow[53].mlp,SeparateRows,qk),d54.withMlp(w,c,Delta,this.#stockNarrow[54].mlp,SeparateRows,qk),a55.withMlp(w,c,Attention,this.#stockNarrow[55].mlp,SeparateRows),
|
|
646
|
+
d56.withMlp(w,c,Delta,this.#stockNarrow[56].mlp,SeparateRows,qk),d57.withMlp(w,c,Delta,this.#stockNarrow[57].mlp,SeparateRows,qk),d58.withMlp(w,c,Delta,this.#stockNarrow[58].mlp,SeparateRows,qk),a59.withMlp(w,c,Attention,this.#stockNarrow[59].mlp,SeparateRows),
|
|
647
|
+
d60.withMlp(w,c,Delta,this.#stockNarrow[60].mlp,SeparateRows,qk),d61.withMlp(w,c,Delta,this.#stockNarrow[61].mlp,SeparateRows,qk),d62.withMlp(w,c,Delta,this.#stockNarrow[62].mlp,SeparateRows,qk),a63.withMlp(w,c,Attention,this.#stockNarrow[63].mlp,SeparateRows),
|
|
648
|
+
];
|
|
649
|
+
this.#knownFiveToEight=[
|
|
650
|
+
d0.withMlp(w,c,Delta,this.#stockFiveToEight[0].mlp,SeparateRows,qk),d1.withMlp(w,c,Delta,this.#stockFiveToEight[1].mlp,SeparateRows,qk),d2.withMlp(w,c,Delta,this.#stockFiveToEight[2].mlp,SeparateRows,qk),a3.withMlp(w,c,Attention,this.#stockFiveToEight[3].mlp,SeparateRows),
|
|
651
|
+
d4.withMlp(w,c,Delta,this.#stockFiveToEight[4].mlp,SeparateRows,qk),d5.withMlp(w,c,Delta,this.#stockFiveToEight[5].mlp,SeparateRows,qk),d6.withMlp(w,c,Delta,this.#stockFiveToEight[6].mlp,SeparateRows,qk),a7.withMlp(w,c,Attention,this.#stockFiveToEight[7].mlp,SeparateRows),
|
|
652
|
+
d8.withMlp(w,c,Delta,this.#stockFiveToEight[8].mlp,SeparateRows,qk),d9.withMlp(w,c,Delta,this.#stockFiveToEight[9].mlp,SeparateRows,qk),d10.withMlp(w,c,Delta,this.#stockFiveToEight[10].mlp,SeparateRows,qk),a11.withMlp(w,c,Attention,this.#stockFiveToEight[11].mlp,SeparateRows),
|
|
653
|
+
d12.withMlp(w,c,Delta,this.#stockFiveToEight[12].mlp,SeparateRows,qk),d13.withMlp(w,c,Delta,this.#stockFiveToEight[13].mlp,SeparateRows,qk),d14.withMlp(w,c,Delta,this.#stockFiveToEight[14].mlp,SeparateRows,qk),a15.withMlp(w,c,Attention,this.#stockFiveToEight[15].mlp,SeparateRows),
|
|
654
|
+
d16.withMlp(w,c,Delta,this.#stockFiveToEight[16].mlp,SeparateRows,qk),d17.withMlp(w,c,Delta,this.#stockFiveToEight[17].mlp,SeparateRows,qk),d18.withMlp(w,c,Delta,this.#stockFiveToEight[18].mlp,SeparateRows,qk),a19.withMlp(w,c,Attention,this.#stockFiveToEight[19].mlp,SeparateRows),
|
|
655
|
+
d20.withMlp(w,c,Delta,this.#stockFiveToEight[20].mlp,SeparateRows,qk),d21.withMlp(w,c,Delta,this.#stockFiveToEight[21].mlp,SeparateRows,qk),d22.withMlp(w,c,Delta,this.#stockFiveToEight[22].mlp,SeparateRows,qk),a23.withMlp(w,c,Attention,this.#stockFiveToEight[23].mlp,SeparateRows),
|
|
656
|
+
d24.withMlp(w,c,Delta,this.#stockFiveToEight[24].mlp,SeparateRows,qk),d25.withMlp(w,c,Delta,this.#stockFiveToEight[25].mlp,SeparateRows,qk),d26.withMlp(w,c,Delta,this.#stockFiveToEight[26].mlp,SeparateRows,qk),a27.withMlp(w,c,Attention,this.#stockFiveToEight[27].mlp,SeparateRows),
|
|
657
|
+
d28.withMlp(w,c,Delta,this.#stockFiveToEight[28].mlp,SeparateRows,qk),d29.withMlp(w,c,Delta,this.#stockFiveToEight[29].mlp,SeparateRows,qk),d30.withMlp(w,c,Delta,this.#stockFiveToEight[30].mlp,SeparateRows,qk),a31.withMlp(w,c,Attention,this.#stockFiveToEight[31].mlp,SeparateRows),
|
|
658
|
+
d32.withMlp(w,c,Delta,this.#stockFiveToEight[32].mlp,SeparateRows,qk),d33.withMlp(w,c,Delta,this.#stockFiveToEight[33].mlp,SeparateRows,qk),d34.withMlp(w,c,Delta,this.#stockFiveToEight[34].mlp,SeparateRows,qk),a35.withMlp(w,c,Attention,this.#stockFiveToEight[35].mlp,SeparateRows),
|
|
659
|
+
d36.withMlp(w,c,Delta,this.#stockFiveToEight[36].mlp,SeparateRows,qk),d37.withMlp(w,c,Delta,this.#stockFiveToEight[37].mlp,SeparateRows,qk),d38.withMlp(w,c,Delta,this.#stockFiveToEight[38].mlp,SeparateRows,qk),a39.withMlp(w,c,Attention,this.#stockFiveToEight[39].mlp,SeparateRows),
|
|
660
|
+
d40.withMlp(w,c,Delta,this.#stockFiveToEight[40].mlp,SeparateRows,qk),d41.withMlp(w,c,Delta,this.#stockFiveToEight[41].mlp,SeparateRows,qk),d42.withMlp(w,c,Delta,this.#stockFiveToEight[42].mlp,SeparateRows,qk),a43.withMlp(w,c,Attention,this.#stockFiveToEight[43].mlp,SeparateRows),
|
|
661
|
+
d44.withMlp(w,c,Delta,this.#stockFiveToEight[44].mlp,SeparateRows,qk),d45.withMlp(w,c,Delta,this.#stockFiveToEight[45].mlp,SeparateRows,qk),d46.withMlp(w,c,Delta,this.#stockFiveToEight[46].mlp,SeparateRows,qk),a47.withMlp(w,c,Attention,this.#stockFiveToEight[47].mlp,SeparateRows),
|
|
662
|
+
d48.withMlp(w,c,Delta,this.#stockFiveToEight[48].mlp,SeparateRows,qk),d49.withMlp(w,c,Delta,this.#stockFiveToEight[49].mlp,SeparateRows,qk),d50.withMlp(w,c,Delta,this.#stockFiveToEight[50].mlp,SeparateRows,qk),a51.withMlp(w,c,Attention,this.#stockFiveToEight[51].mlp,SeparateRows),
|
|
663
|
+
d52.withMlp(w,c,Delta,this.#stockFiveToEight[52].mlp,SeparateRows,qk),d53.withMlp(w,c,Delta,this.#stockFiveToEight[53].mlp,SeparateRows,qk),d54.withMlp(w,c,Delta,this.#stockFiveToEight[54].mlp,SeparateRows,qk),a55.withMlp(w,c,Attention,this.#stockFiveToEight[55].mlp,SeparateRows),
|
|
664
|
+
d56.withMlp(w,c,Delta,this.#stockFiveToEight[56].mlp,SeparateRows,qk),d57.withMlp(w,c,Delta,this.#stockFiveToEight[57].mlp,SeparateRows,qk),d58.withMlp(w,c,Delta,this.#stockFiveToEight[58].mlp,SeparateRows,qk),a59.withMlp(w,c,Attention,this.#stockFiveToEight[59].mlp,SeparateRows),
|
|
665
|
+
d60.withMlp(w,c,Delta,this.#stockFiveToEight[60].mlp,SeparateRows,qk),d61.withMlp(w,c,Delta,this.#stockFiveToEight[61].mlp,SeparateRows,qk),d62.withMlp(w,c,Delta,this.#stockFiveToEight[62].mlp,SeparateRows,qk),a63.withMlp(w,c,Attention,this.#stockFiveToEight[63].mlp,SeparateRows),
|
|
666
|
+
];
|
|
667
|
+
this.#knownWide=[
|
|
668
|
+
new Delta(w,c,d0.prefix,this.#knownNarrow[0].mixer,this.#wide[0].mlp),new Delta(w,c,d1.prefix,this.#knownNarrow[1].mixer,this.#wide[1].mlp),new Delta(w,c,d2.prefix,this.#knownNarrow[2].mixer,this.#wide[2].mlp),new Attention(w,c,a3.prefix,this.#knownNarrow[3].mixer,this.#wide[3].mlp),
|
|
669
|
+
new Delta(w,c,d4.prefix,this.#knownNarrow[4].mixer,this.#wide[4].mlp),new Delta(w,c,d5.prefix,this.#knownNarrow[5].mixer,this.#wide[5].mlp),new Delta(w,c,d6.prefix,this.#knownNarrow[6].mixer,this.#wide[6].mlp),new Attention(w,c,a7.prefix,this.#knownNarrow[7].mixer,this.#wide[7].mlp),
|
|
670
|
+
new Delta(w,c,d8.prefix,this.#knownNarrow[8].mixer,this.#wide[8].mlp),new Delta(w,c,d9.prefix,this.#knownNarrow[9].mixer,this.#wide[9].mlp),new Delta(w,c,d10.prefix,this.#knownNarrow[10].mixer,this.#wide[10].mlp),new Attention(w,c,a11.prefix,this.#knownNarrow[11].mixer,this.#wide[11].mlp),
|
|
671
|
+
new Delta(w,c,d12.prefix,this.#knownNarrow[12].mixer,this.#wide[12].mlp),new Delta(w,c,d13.prefix,this.#knownNarrow[13].mixer,this.#wide[13].mlp),new Delta(w,c,d14.prefix,this.#knownNarrow[14].mixer,this.#wide[14].mlp),new Attention(w,c,a15.prefix,this.#knownNarrow[15].mixer,this.#wide[15].mlp),
|
|
672
|
+
new Delta(w,c,d16.prefix,this.#knownNarrow[16].mixer,this.#wide[16].mlp),new Delta(w,c,d17.prefix,this.#knownNarrow[17].mixer,this.#wide[17].mlp),new Delta(w,c,d18.prefix,this.#knownNarrow[18].mixer,this.#wide[18].mlp),new Attention(w,c,a19.prefix,this.#knownNarrow[19].mixer,this.#wide[19].mlp),
|
|
673
|
+
new Delta(w,c,d20.prefix,this.#knownNarrow[20].mixer,this.#wide[20].mlp),new Delta(w,c,d21.prefix,this.#knownNarrow[21].mixer,this.#wide[21].mlp),new Delta(w,c,d22.prefix,this.#knownNarrow[22].mixer,this.#wide[22].mlp),new Attention(w,c,a23.prefix,this.#knownNarrow[23].mixer,this.#wide[23].mlp),
|
|
674
|
+
new Delta(w,c,d24.prefix,this.#knownNarrow[24].mixer,this.#wide[24].mlp),new Delta(w,c,d25.prefix,this.#knownNarrow[25].mixer,this.#wide[25].mlp),new Delta(w,c,d26.prefix,this.#knownNarrow[26].mixer,this.#wide[26].mlp),new Attention(w,c,a27.prefix,this.#knownNarrow[27].mixer,this.#wide[27].mlp),
|
|
675
|
+
new Delta(w,c,d28.prefix,this.#knownNarrow[28].mixer,this.#wide[28].mlp),new Delta(w,c,d29.prefix,this.#knownNarrow[29].mixer,this.#wide[29].mlp),new Delta(w,c,d30.prefix,this.#knownNarrow[30].mixer,this.#wide[30].mlp),new Attention(w,c,a31.prefix,this.#knownNarrow[31].mixer,this.#wide[31].mlp),
|
|
676
|
+
new Delta(w,c,d32.prefix,this.#knownNarrow[32].mixer,this.#wide[32].mlp),new Delta(w,c,d33.prefix,this.#knownNarrow[33].mixer,this.#wide[33].mlp),new Delta(w,c,d34.prefix,this.#knownNarrow[34].mixer,this.#wide[34].mlp),new Attention(w,c,a35.prefix,this.#knownNarrow[35].mixer,this.#wide[35].mlp),
|
|
677
|
+
new Delta(w,c,d36.prefix,this.#knownNarrow[36].mixer,this.#wide[36].mlp),new Delta(w,c,d37.prefix,this.#knownNarrow[37].mixer,this.#wide[37].mlp),new Delta(w,c,d38.prefix,this.#knownNarrow[38].mixer,this.#wide[38].mlp),new Attention(w,c,a39.prefix,this.#knownNarrow[39].mixer,this.#wide[39].mlp),
|
|
678
|
+
new Delta(w,c,d40.prefix,this.#knownNarrow[40].mixer,this.#wide[40].mlp),new Delta(w,c,d41.prefix,this.#knownNarrow[41].mixer,this.#wide[41].mlp),new Delta(w,c,d42.prefix,this.#knownNarrow[42].mixer,this.#wide[42].mlp),new Attention(w,c,a43.prefix,this.#knownNarrow[43].mixer,this.#wide[43].mlp),
|
|
679
|
+
new Delta(w,c,d44.prefix,this.#knownNarrow[44].mixer,this.#wide[44].mlp),new Delta(w,c,d45.prefix,this.#knownNarrow[45].mixer,this.#wide[45].mlp),new Delta(w,c,d46.prefix,this.#knownNarrow[46].mixer,this.#wide[46].mlp),new Attention(w,c,a47.prefix,this.#knownNarrow[47].mixer,this.#wide[47].mlp),
|
|
680
|
+
new Delta(w,c,d48.prefix,this.#knownNarrow[48].mixer,this.#wide[48].mlp),new Delta(w,c,d49.prefix,this.#knownNarrow[49].mixer,this.#wide[49].mlp),new Delta(w,c,d50.prefix,this.#knownNarrow[50].mixer,this.#wide[50].mlp),new Attention(w,c,a51.prefix,this.#knownNarrow[51].mixer,this.#wide[51].mlp),
|
|
681
|
+
new Delta(w,c,d52.prefix,this.#knownNarrow[52].mixer,this.#wide[52].mlp),new Delta(w,c,d53.prefix,this.#knownNarrow[53].mixer,this.#wide[53].mlp),new Delta(w,c,d54.prefix,this.#knownNarrow[54].mixer,this.#wide[54].mlp),new Attention(w,c,a55.prefix,this.#knownNarrow[55].mixer,this.#wide[55].mlp),
|
|
682
|
+
new Delta(w,c,d56.prefix,this.#knownNarrow[56].mixer,this.#wide[56].mlp),new Delta(w,c,d57.prefix,this.#knownNarrow[57].mixer,this.#wide[57].mlp),new Delta(w,c,d58.prefix,this.#knownNarrow[58].mixer,this.#wide[58].mlp),new Attention(w,c,a59.prefix,this.#knownNarrow[59].mixer,this.#wide[59].mlp),
|
|
683
|
+
new Delta(w,c,d60.prefix,this.#knownNarrow[60].mixer,this.#wide[60].mlp),new Delta(w,c,d61.prefix,this.#knownNarrow[61].mixer,this.#wide[61].mlp),new Delta(w,c,d62.prefix,this.#knownNarrow[62].mixer,this.#wide[62].mlp),new Attention(w,c,a63.prefix,this.#knownNarrow[63].mixer,this.#wide[63].mlp),
|
|
684
|
+
];
|
|
685
|
+
this.#knownTiled=[
|
|
686
|
+
new Delta(w,c,d0.prefix,this.#knownNarrow[0].mixer,this.#tiled[0].mlp),new Delta(w,c,d1.prefix,this.#knownNarrow[1].mixer,this.#tiled[1].mlp),new Delta(w,c,d2.prefix,this.#knownNarrow[2].mixer,this.#tiled[2].mlp),new Attention(w,c,a3.prefix,this.#knownNarrow[3].mixer,this.#tiled[3].mlp),
|
|
687
|
+
new Delta(w,c,d4.prefix,this.#knownNarrow[4].mixer,this.#tiled[4].mlp),new Delta(w,c,d5.prefix,this.#knownNarrow[5].mixer,this.#tiled[5].mlp),new Delta(w,c,d6.prefix,this.#knownNarrow[6].mixer,this.#tiled[6].mlp),new Attention(w,c,a7.prefix,this.#knownNarrow[7].mixer,this.#tiled[7].mlp),
|
|
688
|
+
new Delta(w,c,d8.prefix,this.#knownNarrow[8].mixer,this.#tiled[8].mlp),new Delta(w,c,d9.prefix,this.#knownNarrow[9].mixer,this.#tiled[9].mlp),new Delta(w,c,d10.prefix,this.#knownNarrow[10].mixer,this.#tiled[10].mlp),new Attention(w,c,a11.prefix,this.#knownNarrow[11].mixer,this.#tiled[11].mlp),
|
|
689
|
+
new Delta(w,c,d12.prefix,this.#knownNarrow[12].mixer,this.#tiled[12].mlp),new Delta(w,c,d13.prefix,this.#knownNarrow[13].mixer,this.#tiled[13].mlp),new Delta(w,c,d14.prefix,this.#knownNarrow[14].mixer,this.#tiled[14].mlp),new Attention(w,c,a15.prefix,this.#knownNarrow[15].mixer,this.#tiled[15].mlp),
|
|
690
|
+
new Delta(w,c,d16.prefix,this.#knownNarrow[16].mixer,this.#tiled[16].mlp),new Delta(w,c,d17.prefix,this.#knownNarrow[17].mixer,this.#tiled[17].mlp),new Delta(w,c,d18.prefix,this.#knownNarrow[18].mixer,this.#tiled[18].mlp),new Attention(w,c,a19.prefix,this.#knownNarrow[19].mixer,this.#tiled[19].mlp),
|
|
691
|
+
new Delta(w,c,d20.prefix,this.#knownNarrow[20].mixer,this.#tiled[20].mlp),new Delta(w,c,d21.prefix,this.#knownNarrow[21].mixer,this.#tiled[21].mlp),new Delta(w,c,d22.prefix,this.#knownNarrow[22].mixer,this.#tiled[22].mlp),new Attention(w,c,a23.prefix,this.#knownNarrow[23].mixer,this.#tiled[23].mlp),
|
|
692
|
+
new Delta(w,c,d24.prefix,this.#knownNarrow[24].mixer,this.#tiled[24].mlp),new Delta(w,c,d25.prefix,this.#knownNarrow[25].mixer,this.#tiled[25].mlp),new Delta(w,c,d26.prefix,this.#knownNarrow[26].mixer,this.#tiled[26].mlp),new Attention(w,c,a27.prefix,this.#knownNarrow[27].mixer,this.#tiled[27].mlp),
|
|
693
|
+
new Delta(w,c,d28.prefix,this.#knownNarrow[28].mixer,this.#tiled[28].mlp),new Delta(w,c,d29.prefix,this.#knownNarrow[29].mixer,this.#tiled[29].mlp),new Delta(w,c,d30.prefix,this.#knownNarrow[30].mixer,this.#tiled[30].mlp),new Attention(w,c,a31.prefix,this.#knownNarrow[31].mixer,this.#tiled[31].mlp),
|
|
694
|
+
new Delta(w,c,d32.prefix,this.#knownNarrow[32].mixer,this.#tiled[32].mlp),new Delta(w,c,d33.prefix,this.#knownNarrow[33].mixer,this.#tiled[33].mlp),new Delta(w,c,d34.prefix,this.#knownNarrow[34].mixer,this.#tiled[34].mlp),new Attention(w,c,a35.prefix,this.#knownNarrow[35].mixer,this.#tiled[35].mlp),
|
|
695
|
+
new Delta(w,c,d36.prefix,this.#knownNarrow[36].mixer,this.#tiled[36].mlp),new Delta(w,c,d37.prefix,this.#knownNarrow[37].mixer,this.#tiled[37].mlp),new Delta(w,c,d38.prefix,this.#knownNarrow[38].mixer,this.#tiled[38].mlp),new Attention(w,c,a39.prefix,this.#knownNarrow[39].mixer,this.#tiled[39].mlp),
|
|
696
|
+
new Delta(w,c,d40.prefix,this.#knownNarrow[40].mixer,this.#tiled[40].mlp),new Delta(w,c,d41.prefix,this.#knownNarrow[41].mixer,this.#tiled[41].mlp),new Delta(w,c,d42.prefix,this.#knownNarrow[42].mixer,this.#tiled[42].mlp),new Attention(w,c,a43.prefix,this.#knownNarrow[43].mixer,this.#tiled[43].mlp),
|
|
697
|
+
new Delta(w,c,d44.prefix,this.#knownNarrow[44].mixer,this.#tiled[44].mlp),new Delta(w,c,d45.prefix,this.#knownNarrow[45].mixer,this.#tiled[45].mlp),new Delta(w,c,d46.prefix,this.#knownNarrow[46].mixer,this.#tiled[46].mlp),new Attention(w,c,a47.prefix,this.#knownNarrow[47].mixer,this.#tiled[47].mlp),
|
|
698
|
+
new Delta(w,c,d48.prefix,this.#knownNarrow[48].mixer,this.#tiled[48].mlp),new Delta(w,c,d49.prefix,this.#knownNarrow[49].mixer,this.#tiled[49].mlp),new Delta(w,c,d50.prefix,this.#knownNarrow[50].mixer,this.#tiled[50].mlp),new Attention(w,c,a51.prefix,this.#knownNarrow[51].mixer,this.#tiled[51].mlp),
|
|
699
|
+
new Delta(w,c,d52.prefix,this.#knownNarrow[52].mixer,this.#tiled[52].mlp),new Delta(w,c,d53.prefix,this.#knownNarrow[53].mixer,this.#tiled[53].mlp),new Delta(w,c,d54.prefix,this.#knownNarrow[54].mixer,this.#tiled[54].mlp),new Attention(w,c,a55.prefix,this.#knownNarrow[55].mixer,this.#tiled[55].mlp),
|
|
700
|
+
new Delta(w,c,d56.prefix,this.#knownNarrow[56].mixer,this.#tiled[56].mlp),new Delta(w,c,d57.prefix,this.#knownNarrow[57].mixer,this.#tiled[57].mlp),new Delta(w,c,d58.prefix,this.#knownNarrow[58].mixer,this.#tiled[58].mlp),new Attention(w,c,a59.prefix,this.#knownNarrow[59].mixer,this.#tiled[59].mlp),
|
|
701
|
+
new Delta(w,c,d60.prefix,this.#knownNarrow[60].mixer,this.#tiled[60].mlp),new Delta(w,c,d61.prefix,this.#knownNarrow[61].mixer,this.#tiled[61].mlp),new Delta(w,c,d62.prefix,this.#knownNarrow[62].mixer,this.#tiled[62].mlp),new Attention(w,c,a63.prefix,this.#knownNarrow[63].mixer,this.#tiled[63].mlp),
|
|
702
|
+
];
|
|
703
|
+
this.#knownExpanded=[
|
|
704
|
+
new Delta(w,c,d0.prefix,this.#knownNarrow[0].mixer,this.#expanded[0].mlp),new Delta(w,c,d1.prefix,this.#knownNarrow[1].mixer,this.#expanded[1].mlp),new Delta(w,c,d2.prefix,this.#knownNarrow[2].mixer,this.#expanded[2].mlp),new Attention(w,c,a3.prefix,this.#knownNarrow[3].mixer,this.#expanded[3].mlp),
|
|
705
|
+
new Delta(w,c,d4.prefix,this.#knownNarrow[4].mixer,this.#expanded[4].mlp),new Delta(w,c,d5.prefix,this.#knownNarrow[5].mixer,this.#expanded[5].mlp),new Delta(w,c,d6.prefix,this.#knownNarrow[6].mixer,this.#expanded[6].mlp),new Attention(w,c,a7.prefix,this.#knownNarrow[7].mixer,this.#expanded[7].mlp),
|
|
706
|
+
new Delta(w,c,d8.prefix,this.#knownNarrow[8].mixer,this.#expanded[8].mlp),new Delta(w,c,d9.prefix,this.#knownNarrow[9].mixer,this.#expanded[9].mlp),new Delta(w,c,d10.prefix,this.#knownNarrow[10].mixer,this.#expanded[10].mlp),new Attention(w,c,a11.prefix,this.#knownNarrow[11].mixer,this.#expanded[11].mlp),
|
|
707
|
+
new Delta(w,c,d12.prefix,this.#knownNarrow[12].mixer,this.#expanded[12].mlp),new Delta(w,c,d13.prefix,this.#knownNarrow[13].mixer,this.#expanded[13].mlp),new Delta(w,c,d14.prefix,this.#knownNarrow[14].mixer,this.#expanded[14].mlp),new Attention(w,c,a15.prefix,this.#knownNarrow[15].mixer,this.#expanded[15].mlp),
|
|
708
|
+
new Delta(w,c,d16.prefix,this.#knownNarrow[16].mixer,this.#expanded[16].mlp),new Delta(w,c,d17.prefix,this.#knownNarrow[17].mixer,this.#expanded[17].mlp),new Delta(w,c,d18.prefix,this.#knownNarrow[18].mixer,this.#expanded[18].mlp),new Attention(w,c,a19.prefix,this.#knownNarrow[19].mixer,this.#expanded[19].mlp),
|
|
709
|
+
new Delta(w,c,d20.prefix,this.#knownNarrow[20].mixer,this.#expanded[20].mlp),new Delta(w,c,d21.prefix,this.#knownNarrow[21].mixer,this.#expanded[21].mlp),new Delta(w,c,d22.prefix,this.#knownNarrow[22].mixer,this.#expanded[22].mlp),new Attention(w,c,a23.prefix,this.#knownNarrow[23].mixer,this.#expanded[23].mlp),
|
|
710
|
+
new Delta(w,c,d24.prefix,this.#knownNarrow[24].mixer,this.#expanded[24].mlp),new Delta(w,c,d25.prefix,this.#knownNarrow[25].mixer,this.#expanded[25].mlp),new Delta(w,c,d26.prefix,this.#knownNarrow[26].mixer,this.#expanded[26].mlp),new Attention(w,c,a27.prefix,this.#knownNarrow[27].mixer,this.#expanded[27].mlp),
|
|
711
|
+
new Delta(w,c,d28.prefix,this.#knownNarrow[28].mixer,this.#expanded[28].mlp),new Delta(w,c,d29.prefix,this.#knownNarrow[29].mixer,this.#expanded[29].mlp),new Delta(w,c,d30.prefix,this.#knownNarrow[30].mixer,this.#expanded[30].mlp),new Attention(w,c,a31.prefix,this.#knownNarrow[31].mixer,this.#expanded[31].mlp),
|
|
712
|
+
new Delta(w,c,d32.prefix,this.#knownNarrow[32].mixer,this.#expanded[32].mlp),new Delta(w,c,d33.prefix,this.#knownNarrow[33].mixer,this.#expanded[33].mlp),new Delta(w,c,d34.prefix,this.#knownNarrow[34].mixer,this.#expanded[34].mlp),new Attention(w,c,a35.prefix,this.#knownNarrow[35].mixer,this.#expanded[35].mlp),
|
|
713
|
+
new Delta(w,c,d36.prefix,this.#knownNarrow[36].mixer,this.#expanded[36].mlp),new Delta(w,c,d37.prefix,this.#knownNarrow[37].mixer,this.#expanded[37].mlp),new Delta(w,c,d38.prefix,this.#knownNarrow[38].mixer,this.#expanded[38].mlp),new Attention(w,c,a39.prefix,this.#knownNarrow[39].mixer,this.#expanded[39].mlp),
|
|
714
|
+
new Delta(w,c,d40.prefix,this.#knownNarrow[40].mixer,this.#expanded[40].mlp),new Delta(w,c,d41.prefix,this.#knownNarrow[41].mixer,this.#expanded[41].mlp),new Delta(w,c,d42.prefix,this.#knownNarrow[42].mixer,this.#expanded[42].mlp),new Attention(w,c,a43.prefix,this.#knownNarrow[43].mixer,this.#expanded[43].mlp),
|
|
715
|
+
new Delta(w,c,d44.prefix,this.#knownNarrow[44].mixer,this.#expanded[44].mlp),new Delta(w,c,d45.prefix,this.#knownNarrow[45].mixer,this.#expanded[45].mlp),new Delta(w,c,d46.prefix,this.#knownNarrow[46].mixer,this.#expanded[46].mlp),new Attention(w,c,a47.prefix,this.#knownNarrow[47].mixer,this.#expanded[47].mlp),
|
|
716
|
+
new Delta(w,c,d48.prefix,this.#knownNarrow[48].mixer,this.#expanded[48].mlp),new Delta(w,c,d49.prefix,this.#knownNarrow[49].mixer,this.#expanded[49].mlp),new Delta(w,c,d50.prefix,this.#knownNarrow[50].mixer,this.#expanded[50].mlp),new Attention(w,c,a51.prefix,this.#knownNarrow[51].mixer,this.#expanded[51].mlp),
|
|
717
|
+
new Delta(w,c,d52.prefix,this.#knownNarrow[52].mixer,this.#expanded[52].mlp),new Delta(w,c,d53.prefix,this.#knownNarrow[53].mixer,this.#expanded[53].mlp),new Delta(w,c,d54.prefix,this.#knownNarrow[54].mixer,this.#expanded[54].mlp),new Attention(w,c,a55.prefix,this.#knownNarrow[55].mixer,this.#expanded[55].mlp),
|
|
718
|
+
new Delta(w,c,d56.prefix,this.#knownNarrow[56].mixer,this.#expanded[56].mlp),new Delta(w,c,d57.prefix,this.#knownNarrow[57].mixer,this.#expanded[57].mlp),new Delta(w,c,d58.prefix,this.#knownNarrow[58].mixer,this.#expanded[58].mlp),new Attention(w,c,a59.prefix,this.#knownNarrow[59].mixer,this.#expanded[59].mlp),
|
|
719
|
+
new Delta(w,c,d60.prefix,this.#knownNarrow[60].mixer,this.#expanded[60].mlp),new Delta(w,c,d61.prefix,this.#knownNarrow[61].mixer,this.#expanded[61].mlp),new Delta(w,c,d62.prefix,this.#knownNarrow[62].mixer,this.#expanded[62].mlp),new Attention(w,c,a63.prefix,this.#knownNarrow[63].mixer,this.#expanded[63].mlp),
|
|
720
|
+
];
|
|
721
|
+
const mediaNarrow:Layers=[
|
|
722
|
+
this.#stockNarrow[0],this.#stockNarrow[1],this.#stockNarrow[2],a3.mediaLayer(w,c,Attention,this.#stockNarrow[3].mlp,Plain),
|
|
723
|
+
this.#stockNarrow[4],this.#stockNarrow[5],this.#stockNarrow[6],a7.mediaLayer(w,c,Attention,this.#stockNarrow[7].mlp,Plain),
|
|
724
|
+
this.#stockNarrow[8],this.#stockNarrow[9],this.#stockNarrow[10],a11.mediaLayer(w,c,Attention,this.#stockNarrow[11].mlp,Plain),
|
|
725
|
+
this.#stockNarrow[12],this.#stockNarrow[13],this.#stockNarrow[14],a15.mediaLayer(w,c,Attention,this.#stockNarrow[15].mlp,Plain),
|
|
726
|
+
this.#stockNarrow[16],this.#stockNarrow[17],this.#stockNarrow[18],a19.mediaLayer(w,c,Attention,this.#stockNarrow[19].mlp,Plain),
|
|
727
|
+
this.#stockNarrow[20],this.#stockNarrow[21],this.#stockNarrow[22],a23.mediaLayer(w,c,Attention,this.#stockNarrow[23].mlp,Plain),
|
|
728
|
+
this.#stockNarrow[24],this.#stockNarrow[25],this.#stockNarrow[26],a27.mediaLayer(w,c,Attention,this.#stockNarrow[27].mlp,Plain),
|
|
729
|
+
this.#stockNarrow[28],this.#stockNarrow[29],this.#stockNarrow[30],a31.mediaLayer(w,c,Attention,this.#stockNarrow[31].mlp,Plain),
|
|
730
|
+
this.#stockNarrow[32],this.#stockNarrow[33],this.#stockNarrow[34],a35.mediaLayer(w,c,Attention,this.#stockNarrow[35].mlp,Plain),
|
|
731
|
+
this.#stockNarrow[36],this.#stockNarrow[37],this.#stockNarrow[38],a39.mediaLayer(w,c,Attention,this.#stockNarrow[39].mlp,Plain),
|
|
732
|
+
this.#stockNarrow[40],this.#stockNarrow[41],this.#stockNarrow[42],a43.mediaLayer(w,c,Attention,this.#stockNarrow[43].mlp,Plain),
|
|
733
|
+
this.#stockNarrow[44],this.#stockNarrow[45],this.#stockNarrow[46],a47.mediaLayer(w,c,Attention,this.#stockNarrow[47].mlp,Plain),
|
|
734
|
+
this.#stockNarrow[48],this.#stockNarrow[49],this.#stockNarrow[50],a51.mediaLayer(w,c,Attention,this.#stockNarrow[51].mlp,Plain),
|
|
735
|
+
this.#stockNarrow[52],this.#stockNarrow[53],this.#stockNarrow[54],a55.mediaLayer(w,c,Attention,this.#stockNarrow[55].mlp,Plain),
|
|
736
|
+
this.#stockNarrow[56],this.#stockNarrow[57],this.#stockNarrow[58],a59.mediaLayer(w,c,Attention,this.#stockNarrow[59].mlp,Plain),
|
|
737
|
+
this.#stockNarrow[60],this.#stockNarrow[61],this.#stockNarrow[62],a63.mediaLayer(w,c,Attention,this.#stockNarrow[63].mlp,Plain),
|
|
738
|
+
];
|
|
739
|
+
const mediaFiveToEight:Layers=[
|
|
740
|
+
this.#stockFiveToEight[0],this.#stockFiveToEight[1],this.#stockFiveToEight[2],a3.mediaLayer(w,c,Attention,this.#stockFiveToEight[3].mlp,Plain),
|
|
741
|
+
this.#stockFiveToEight[4],this.#stockFiveToEight[5],this.#stockFiveToEight[6],a7.mediaLayer(w,c,Attention,this.#stockFiveToEight[7].mlp,Plain),
|
|
742
|
+
this.#stockFiveToEight[8],this.#stockFiveToEight[9],this.#stockFiveToEight[10],a11.mediaLayer(w,c,Attention,this.#stockFiveToEight[11].mlp,Plain),
|
|
743
|
+
this.#stockFiveToEight[12],this.#stockFiveToEight[13],this.#stockFiveToEight[14],a15.mediaLayer(w,c,Attention,this.#stockFiveToEight[15].mlp,Plain),
|
|
744
|
+
this.#stockFiveToEight[16],this.#stockFiveToEight[17],this.#stockFiveToEight[18],a19.mediaLayer(w,c,Attention,this.#stockFiveToEight[19].mlp,Plain),
|
|
745
|
+
this.#stockFiveToEight[20],this.#stockFiveToEight[21],this.#stockFiveToEight[22],a23.mediaLayer(w,c,Attention,this.#stockFiveToEight[23].mlp,Plain),
|
|
746
|
+
this.#stockFiveToEight[24],this.#stockFiveToEight[25],this.#stockFiveToEight[26],a27.mediaLayer(w,c,Attention,this.#stockFiveToEight[27].mlp,Plain),
|
|
747
|
+
this.#stockFiveToEight[28],this.#stockFiveToEight[29],this.#stockFiveToEight[30],a31.mediaLayer(w,c,Attention,this.#stockFiveToEight[31].mlp,Plain),
|
|
748
|
+
this.#stockFiveToEight[32],this.#stockFiveToEight[33],this.#stockFiveToEight[34],a35.mediaLayer(w,c,Attention,this.#stockFiveToEight[35].mlp,Plain),
|
|
749
|
+
this.#stockFiveToEight[36],this.#stockFiveToEight[37],this.#stockFiveToEight[38],a39.mediaLayer(w,c,Attention,this.#stockFiveToEight[39].mlp,Plain),
|
|
750
|
+
this.#stockFiveToEight[40],this.#stockFiveToEight[41],this.#stockFiveToEight[42],a43.mediaLayer(w,c,Attention,this.#stockFiveToEight[43].mlp,Plain),
|
|
751
|
+
this.#stockFiveToEight[44],this.#stockFiveToEight[45],this.#stockFiveToEight[46],a47.mediaLayer(w,c,Attention,this.#stockFiveToEight[47].mlp,Plain),
|
|
752
|
+
this.#stockFiveToEight[48],this.#stockFiveToEight[49],this.#stockFiveToEight[50],a51.mediaLayer(w,c,Attention,this.#stockFiveToEight[51].mlp,Plain),
|
|
753
|
+
this.#stockFiveToEight[52],this.#stockFiveToEight[53],this.#stockFiveToEight[54],a55.mediaLayer(w,c,Attention,this.#stockFiveToEight[55].mlp,Plain),
|
|
754
|
+
this.#stockFiveToEight[56],this.#stockFiveToEight[57],this.#stockFiveToEight[58],a59.mediaLayer(w,c,Attention,this.#stockFiveToEight[59].mlp,Plain),
|
|
755
|
+
this.#stockFiveToEight[60],this.#stockFiveToEight[61],this.#stockFiveToEight[62],a63.mediaLayer(w,c,Attention,this.#stockFiveToEight[63].mlp,Plain),
|
|
756
|
+
];
|
|
757
|
+
const mediaWide:Layers=[
|
|
758
|
+
this.#wide[0],this.#wide[1],this.#wide[2],a3.mediaLayer(w,c,Attention,this.#wide[3].mlp,Plain),
|
|
759
|
+
this.#wide[4],this.#wide[5],this.#wide[6],a7.mediaLayer(w,c,Attention,this.#wide[7].mlp,Plain),
|
|
760
|
+
this.#wide[8],this.#wide[9],this.#wide[10],a11.mediaLayer(w,c,Attention,this.#wide[11].mlp,Plain),
|
|
761
|
+
this.#wide[12],this.#wide[13],this.#wide[14],a15.mediaLayer(w,c,Attention,this.#wide[15].mlp,Plain),
|
|
762
|
+
this.#wide[16],this.#wide[17],this.#wide[18],a19.mediaLayer(w,c,Attention,this.#wide[19].mlp,Plain),
|
|
763
|
+
this.#wide[20],this.#wide[21],this.#wide[22],a23.mediaLayer(w,c,Attention,this.#wide[23].mlp,Plain),
|
|
764
|
+
this.#wide[24],this.#wide[25],this.#wide[26],a27.mediaLayer(w,c,Attention,this.#wide[27].mlp,Plain),
|
|
765
|
+
this.#wide[28],this.#wide[29],this.#wide[30],a31.mediaLayer(w,c,Attention,this.#wide[31].mlp,Plain),
|
|
766
|
+
this.#wide[32],this.#wide[33],this.#wide[34],a35.mediaLayer(w,c,Attention,this.#wide[35].mlp,Plain),
|
|
767
|
+
this.#wide[36],this.#wide[37],this.#wide[38],a39.mediaLayer(w,c,Attention,this.#wide[39].mlp,Plain),
|
|
768
|
+
this.#wide[40],this.#wide[41],this.#wide[42],a43.mediaLayer(w,c,Attention,this.#wide[43].mlp,Plain),
|
|
769
|
+
this.#wide[44],this.#wide[45],this.#wide[46],a47.mediaLayer(w,c,Attention,this.#wide[47].mlp,Plain),
|
|
770
|
+
this.#wide[48],this.#wide[49],this.#wide[50],a51.mediaLayer(w,c,Attention,this.#wide[51].mlp,Plain),
|
|
771
|
+
this.#wide[52],this.#wide[53],this.#wide[54],a55.mediaLayer(w,c,Attention,this.#wide[55].mlp,Plain),
|
|
772
|
+
this.#wide[56],this.#wide[57],this.#wide[58],a59.mediaLayer(w,c,Attention,this.#wide[59].mlp,Plain),
|
|
773
|
+
this.#wide[60],this.#wide[61],this.#wide[62],a63.mediaLayer(w,c,Attention,this.#wide[63].mlp,Plain),
|
|
774
|
+
];
|
|
775
|
+
const mediaTiled:Layers=[
|
|
776
|
+
this.#tiled[0],this.#tiled[1],this.#tiled[2],a3.mediaLayer(w,c,Attention,this.#tiled[3].mlp,Plain),
|
|
777
|
+
this.#tiled[4],this.#tiled[5],this.#tiled[6],a7.mediaLayer(w,c,Attention,this.#tiled[7].mlp,Plain),
|
|
778
|
+
this.#tiled[8],this.#tiled[9],this.#tiled[10],a11.mediaLayer(w,c,Attention,this.#tiled[11].mlp,Plain),
|
|
779
|
+
this.#tiled[12],this.#tiled[13],this.#tiled[14],a15.mediaLayer(w,c,Attention,this.#tiled[15].mlp,Plain),
|
|
780
|
+
this.#tiled[16],this.#tiled[17],this.#tiled[18],a19.mediaLayer(w,c,Attention,this.#tiled[19].mlp,Plain),
|
|
781
|
+
this.#tiled[20],this.#tiled[21],this.#tiled[22],a23.mediaLayer(w,c,Attention,this.#tiled[23].mlp,Plain),
|
|
782
|
+
this.#tiled[24],this.#tiled[25],this.#tiled[26],a27.mediaLayer(w,c,Attention,this.#tiled[27].mlp,Plain),
|
|
783
|
+
this.#tiled[28],this.#tiled[29],this.#tiled[30],a31.mediaLayer(w,c,Attention,this.#tiled[31].mlp,Plain),
|
|
784
|
+
this.#tiled[32],this.#tiled[33],this.#tiled[34],a35.mediaLayer(w,c,Attention,this.#tiled[35].mlp,Plain),
|
|
785
|
+
this.#tiled[36],this.#tiled[37],this.#tiled[38],a39.mediaLayer(w,c,Attention,this.#tiled[39].mlp,Plain),
|
|
786
|
+
this.#tiled[40],this.#tiled[41],this.#tiled[42],a43.mediaLayer(w,c,Attention,this.#tiled[43].mlp,Plain),
|
|
787
|
+
this.#tiled[44],this.#tiled[45],this.#tiled[46],a47.mediaLayer(w,c,Attention,this.#tiled[47].mlp,Plain),
|
|
788
|
+
this.#tiled[48],this.#tiled[49],this.#tiled[50],a51.mediaLayer(w,c,Attention,this.#tiled[51].mlp,Plain),
|
|
789
|
+
this.#tiled[52],this.#tiled[53],this.#tiled[54],a55.mediaLayer(w,c,Attention,this.#tiled[55].mlp,Plain),
|
|
790
|
+
this.#tiled[56],this.#tiled[57],this.#tiled[58],a59.mediaLayer(w,c,Attention,this.#tiled[59].mlp,Plain),
|
|
791
|
+
this.#tiled[60],this.#tiled[61],this.#tiled[62],a63.mediaLayer(w,c,Attention,this.#tiled[63].mlp,Plain),
|
|
792
|
+
];
|
|
793
|
+
const mediaExpanded:Layers=[
|
|
794
|
+
this.#expanded[0],this.#expanded[1],this.#expanded[2],a3.mediaLayer(w,c,Attention,this.#expanded[3].mlp,Plain),
|
|
795
|
+
this.#expanded[4],this.#expanded[5],this.#expanded[6],a7.mediaLayer(w,c,Attention,this.#expanded[7].mlp,Plain),
|
|
796
|
+
this.#expanded[8],this.#expanded[9],this.#expanded[10],a11.mediaLayer(w,c,Attention,this.#expanded[11].mlp,Plain),
|
|
797
|
+
this.#expanded[12],this.#expanded[13],this.#expanded[14],a15.mediaLayer(w,c,Attention,this.#expanded[15].mlp,Plain),
|
|
798
|
+
this.#expanded[16],this.#expanded[17],this.#expanded[18],a19.mediaLayer(w,c,Attention,this.#expanded[19].mlp,Plain),
|
|
799
|
+
this.#expanded[20],this.#expanded[21],this.#expanded[22],a23.mediaLayer(w,c,Attention,this.#expanded[23].mlp,Plain),
|
|
800
|
+
this.#expanded[24],this.#expanded[25],this.#expanded[26],a27.mediaLayer(w,c,Attention,this.#expanded[27].mlp,Plain),
|
|
801
|
+
this.#expanded[28],this.#expanded[29],this.#expanded[30],a31.mediaLayer(w,c,Attention,this.#expanded[31].mlp,Plain),
|
|
802
|
+
this.#expanded[32],this.#expanded[33],this.#expanded[34],a35.mediaLayer(w,c,Attention,this.#expanded[35].mlp,Plain),
|
|
803
|
+
this.#expanded[36],this.#expanded[37],this.#expanded[38],a39.mediaLayer(w,c,Attention,this.#expanded[39].mlp,Plain),
|
|
804
|
+
this.#expanded[40],this.#expanded[41],this.#expanded[42],a43.mediaLayer(w,c,Attention,this.#expanded[43].mlp,Plain),
|
|
805
|
+
this.#expanded[44],this.#expanded[45],this.#expanded[46],a47.mediaLayer(w,c,Attention,this.#expanded[47].mlp,Plain),
|
|
806
|
+
this.#expanded[48],this.#expanded[49],this.#expanded[50],a51.mediaLayer(w,c,Attention,this.#expanded[51].mlp,Plain),
|
|
807
|
+
this.#expanded[52],this.#expanded[53],this.#expanded[54],a55.mediaLayer(w,c,Attention,this.#expanded[55].mlp,Plain),
|
|
808
|
+
this.#expanded[56],this.#expanded[57],this.#expanded[58],a59.mediaLayer(w,c,Attention,this.#expanded[59].mlp,Plain),
|
|
809
|
+
this.#expanded[60],this.#expanded[61],this.#expanded[62],a63.mediaLayer(w,c,Attention,this.#expanded[63].mlp,Plain),
|
|
810
|
+
];
|
|
811
|
+
this.#targets=new Map([
|
|
812
|
+
...d0.targets(),...d1.targets(),...d2.targets(),...a3.targets(),...d4.targets(),...d5.targets(),...d6.targets(),...a7.targets(),
|
|
813
|
+
...d8.targets(),...d9.targets(),...d10.targets(),...a11.targets(),...d12.targets(),...d13.targets(),...d14.targets(),...a15.targets(),
|
|
814
|
+
...d16.targets(),...d17.targets(),...d18.targets(),...a19.targets(),...d20.targets(),...d21.targets(),...d22.targets(),...a23.targets(),
|
|
815
|
+
...d24.targets(),...d25.targets(),...d26.targets(),...a27.targets(),...d28.targets(),...d29.targets(),...d30.targets(),...a31.targets(),
|
|
816
|
+
...d32.targets(),...d33.targets(),...d34.targets(),...a35.targets(),...d36.targets(),...d37.targets(),...d38.targets(),...a39.targets(),
|
|
817
|
+
...d40.targets(),...d41.targets(),...d42.targets(),...a43.targets(),...d44.targets(),...d45.targets(),...d46.targets(),...a47.targets(),
|
|
818
|
+
...d48.targets(),...d49.targets(),...d50.targets(),...a51.targets(),...d52.targets(),...d53.targets(),...d54.targets(),...a55.targets(),
|
|
819
|
+
...d56.targets(),...d57.targets(),...d58.targets(),...a59.targets(),...d60.targets(),...d61.targets(),...d62.targets(),...a63.targets(),
|
|
820
|
+
]);
|
|
821
|
+
const oneDecode:RowsForward=(ids,state,capture)=>this.#submitted(this.#one,ids,state,"decode",capture);
|
|
822
|
+
const rowsDecode:RowsForward=(ids,state,capture)=>this.#submitted(this.#rows,ids,state,"decode",capture);
|
|
823
|
+
const fourDecode:RowsForward=(ids,state,capture)=>this.#submitted(this.#four,ids,state,"decode",capture);
|
|
824
|
+
const mmaDecode:RowsForward=(ids,state,capture)=>this.#submitted(this.#mma,ids,state,"decode",capture);
|
|
825
|
+
const wideDecode:RowsForward=(ids,state,capture)=>this.#lazy(this.#wide,ids,state,"decode",capture);
|
|
826
|
+
const tiledDecode:RowsForward=(ids,state,capture)=>this.#lazy(this.#tiled,ids,state,"decode",capture);
|
|
827
|
+
const expandedDecode:RowsForward=(ids,state,capture)=>this.#lazy(this.#expanded,ids,state,"decode",capture);
|
|
828
|
+
const oneWindow:RowsForward=(ids,state,capture)=>this.#submitted(this.#one,ids,state,"window",capture);
|
|
829
|
+
const rowsWindow:RowsForward=(ids,state,capture)=>this.#submitted(this.#rows,ids,state,"window",capture);
|
|
830
|
+
const fourWindow:RowsForward=(ids,state,capture)=>this.#submitted(this.#four,ids,state,"window",capture);
|
|
831
|
+
const mmaWindow:RowsForward=(ids,state,capture)=>this.#submitted(this.#mma,ids,state,"window",capture);
|
|
832
|
+
const wideWindow:RowsForward=(ids,state,capture)=>this.#bounded(this.#wide,ids,state,"window",capture);
|
|
833
|
+
const tiledWindow:RowsForward=(ids,state,capture)=>this.#bounded(this.#tiled,ids,state,"window",capture);
|
|
834
|
+
const expandedWindow:RowsForward=(ids,state,capture)=>this.#bounded(this.#expanded,ids,state,"window",capture);
|
|
835
|
+
this.decode=Object.freeze({
|
|
836
|
+
1:oneDecode,2:rowsDecode,3:rowsDecode,4:fourDecode,5:mmaDecode,6:mmaDecode,7:mmaDecode,8:mmaDecode,
|
|
837
|
+
9:wideDecode,10:wideDecode,11:wideDecode,12:wideDecode,13:wideDecode,14:wideDecode,15:wideDecode,
|
|
838
|
+
16:tiledDecode,17:tiledDecode,18:tiledDecode,19:tiledDecode,20:tiledDecode,21:tiledDecode,22:tiledDecode,23:tiledDecode,
|
|
839
|
+
24:tiledDecode,25:tiledDecode,26:tiledDecode,27:tiledDecode,28:tiledDecode,29:tiledDecode,30:tiledDecode,31:tiledDecode,32:tiledDecode,
|
|
840
|
+
|
|
841
|
+
default: expandedDecode,
|
|
842
|
+
});
|
|
843
|
+
this.verify=Object.freeze({
|
|
844
|
+
1:oneWindow,2:rowsWindow,3:rowsWindow,4:fourWindow,5:mmaWindow,6:mmaWindow,7:mmaWindow,8:mmaWindow,
|
|
845
|
+
9:wideWindow,10:wideWindow,11:wideWindow,12:wideWindow,13:wideWindow,14:wideWindow,15:wideWindow,
|
|
846
|
+
16:tiledWindow,17:tiledWindow,18:tiledWindow,19:tiledWindow,20:tiledWindow,21:tiledWindow,22:tiledWindow,23:tiledWindow,
|
|
847
|
+
24:tiledWindow,25:tiledWindow,26:tiledWindow,27:tiledWindow,28:tiledWindow,29:tiledWindow,30:tiledWindow,31:tiledWindow,32:tiledWindow,
|
|
848
|
+
33:expandedWindow,34:expandedWindow,35:expandedWindow,36:expandedWindow,37:expandedWindow,38:expandedWindow,39:expandedWindow,40:expandedWindow,
|
|
849
|
+
41:expandedWindow,42:expandedWindow,43:expandedWindow,44:expandedWindow,45:expandedWindow,46:expandedWindow,47:expandedWindow,48:expandedWindow,
|
|
850
|
+
49:expandedWindow,50:expandedWindow,51:expandedWindow,52:expandedWindow,53:expandedWindow,54:expandedWindow,55:expandedWindow,56:expandedWindow,
|
|
851
|
+
57:expandedWindow,58:expandedWindow,59:expandedWindow,60:expandedWindow,61:expandedWindow,62:expandedWindow,63:expandedWindow,64:expandedWindow,
|
|
852
|
+
|
|
853
|
+
default: expandedWindow,
|
|
854
|
+
});
|
|
855
|
+
this.prompt=[
|
|
856
|
+
{from:1,to:1,operation:oneWindow},{from:2,to:3,operation:rowsWindow},{from:4,to:4,operation:fourWindow},
|
|
857
|
+
{from:5,to:8,operation:mmaWindow},{from:9,to:15,operation:wideWindow},{from:16,to:32,operation:tiledWindow},
|
|
858
|
+
{from:33,to:639,operation:expandedWindow},
|
|
859
|
+
{from:640,to:1024,unit:"ane",operation:(ids,state,capture)=>this.#bounded(this.#ane1024,ids,state,"window",capture)},
|
|
860
|
+
{from:1025,to:1536,unit:"ane",operation:(ids,state,capture)=>this.#bounded(this.#ane1536,ids,state,"window",capture)},
|
|
861
|
+
{from:1537,to:2048,unit:"ane",operation:(ids,state,capture)=>this.#bounded(this.#ane2048,ids,state,"window",capture)},
|
|
862
|
+
{from:2049,to:Infinity,operation:expandedWindow},
|
|
863
|
+
];
|
|
864
|
+
const chunkNarrow:Layers=[
|
|
865
|
+
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),
|
|
866
|
+
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),
|
|
867
|
+
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),
|
|
868
|
+
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),
|
|
869
|
+
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),
|
|
870
|
+
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),
|
|
871
|
+
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),
|
|
872
|
+
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),
|
|
873
|
+
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),
|
|
874
|
+
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),
|
|
875
|
+
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),
|
|
876
|
+
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),
|
|
877
|
+
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),
|
|
878
|
+
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),
|
|
879
|
+
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),
|
|
880
|
+
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),
|
|
881
|
+
];
|
|
882
|
+
this.multiSequencePrompt=[
|
|
883
|
+
{from:1,to:1,operation:oneWindow},{from:2,to:3,operation:rowsWindow},{from:4,to:4,operation:fourWindow},
|
|
884
|
+
{from:5,to:8,operation:mmaWindow},
|
|
885
|
+
{from:9,to:15,operation:wideWindow},{from:16,to:32,operation:tiledWindow},{from:33,to:Infinity,operation:expandedWindow},
|
|
886
|
+
];
|
|
887
|
+
const narrowChunk:RowsForward=(ids,state,capture)=>this.#bounded(chunkNarrow,ids,state,"window",capture);
|
|
888
|
+
const fiveChunk:RowsForward=(ids,state,capture)=>this.#bounded(this.#stockFiveToEight,ids,state,"window",capture);
|
|
889
|
+
this.multiSequencePromptChunk=[
|
|
890
|
+
{from:1,to:4,operation:narrowChunk},{from:5,to:8,operation:fiveChunk},
|
|
891
|
+
{from:9,to:15,operation:wideWindow},{from:16,to:32,operation:tiledWindow},{from:33,to:Infinity,operation:expandedWindow},
|
|
892
|
+
];
|
|
893
|
+
this.promptChunk=[
|
|
894
|
+
{from:1,to:4,operation:narrowChunk},{from:5,to:8,operation:fiveChunk},
|
|
895
|
+
{from:9,to:15,operation:wideWindow},{from:16,to:32,operation:tiledWindow},{from:33,to:639,operation:expandedWindow},
|
|
896
|
+
{from:640,to:1024,unit:"ane",operation:(ids,state,capture)=>this.#bounded(this.#ane1024,ids,state,"window",capture)},
|
|
897
|
+
{from:1025,to:1536,unit:"ane",operation:(ids,state,capture)=>this.#bounded(this.#ane1536,ids,state,"window",capture)},
|
|
898
|
+
{from:1537,to:2048,unit:"ane",operation:(ids,state,capture)=>this.#bounded(this.#ane2048,ids,state,"window",capture)},
|
|
899
|
+
{from:2049,to:Infinity,operation:expandedWindow},
|
|
900
|
+
];
|
|
901
|
+
const knownNarrow:MlxPhaseForward=(ids,state)=>this.#lazy(this.#knownNarrow,ids,state,"committed",null);
|
|
902
|
+
const knownFiveToEight:MlxPhaseForward=(ids,state)=>this.#bounded(this.#knownFiveToEight,ids,state,"committed",null);
|
|
903
|
+
const knownWide:MlxPhaseForward=(ids,state)=>this.#bounded(this.#knownWide,ids,state,"committed",null);
|
|
904
|
+
const knownTiled:MlxPhaseForward=(ids,state)=>this.#bounded(this.#knownTiled,ids,state,"committed",null);
|
|
905
|
+
const knownExpanded:MlxPhaseForward=(ids,state)=>this.#bounded(this.#knownExpanded,ids,state,"committed",null);
|
|
906
|
+
this.appendKnown=[{from:1,to:4,operation:knownNarrow},{from:5,to:8,operation:knownFiveToEight},
|
|
907
|
+
{from:9,to:15,operation:knownWide},{from:16,to:32,operation:knownTiled},{from:33,to:Infinity,operation:knownExpanded}];
|
|
908
|
+
const headRows=new Affine4Rows(this.lmHead),headMma=new Affine4Mma(this.lmHead);
|
|
909
|
+
const plainHead:MlxHeadForward=hidden=>this.#head.forward(hidden),rowsHead:MlxHeadForward=hidden=>headRows.forward(hidden),mmaHead:MlxHeadForward=hidden=>headMma.forward(hidden);
|
|
910
|
+
this.logits=Object.freeze({
|
|
911
|
+
1:plainHead,2:rowsHead,3:rowsHead,4:rowsHead,5:mmaHead,6:mmaHead,7:mmaHead,8:mmaHead,
|
|
912
|
+
9:plainHead,10:plainHead,11:plainHead,12:plainHead,13:plainHead,14:plainHead,15:plainHead,16:plainHead,
|
|
913
|
+
17:plainHead,18:plainHead,19:plainHead,20:plainHead,21:plainHead,22:plainHead,23:plainHead,24:plainHead,
|
|
914
|
+
25:plainHead,26:plainHead,27:plainHead,28:plainHead,29:plainHead,30:plainHead,31:plainHead,32:plainHead,
|
|
915
|
+
33:plainHead,34:plainHead,35:plainHead,36:plainHead,37:plainHead,38:plainHead,39:plainHead,40:plainHead,
|
|
916
|
+
41:plainHead,42:plainHead,43:plainHead,44:plainHead,45:plainHead,46:plainHead,47:plainHead,48:plainHead,
|
|
917
|
+
49:plainHead,50:plainHead,51:plainHead,52:plainHead,53:plainHead,54:plainHead,55:plainHead,56:plainHead,
|
|
918
|
+
57:plainHead,58:plainHead,59:plainHead,60:plainHead,61:plainHead,62:plainHead,63:plainHead,64:plainHead,
|
|
919
|
+
|
|
920
|
+
default: plainHead,
|
|
921
|
+
});
|
|
922
|
+
this.sequence=[
|
|
923
|
+
{from:1,to:1,operation:(ids,state)=>oneWindow(ids,state,null)},
|
|
924
|
+
{from:2,to:3,operation:(ids,state)=>rowsWindow(ids,state,null)},
|
|
925
|
+
{from:4,to:4,operation:(ids,state)=>fourWindow(ids,state,null)},
|
|
926
|
+
{from:5,to:8,operation:(ids,state)=>mmaWindow(ids,state,null)},
|
|
927
|
+
{from:9,to:15,operation:(ids,state)=>wideWindow(ids,state,null)},
|
|
928
|
+
{from:16,to:32,operation:(ids,state)=>tiledWindow(ids,state,null)},
|
|
929
|
+
{from:33,to:Infinity,operation:(ids,state)=>expandedWindow(ids,state,null)},
|
|
930
|
+
];
|
|
931
|
+
this.mediaPrefill=[
|
|
932
|
+
{from:1,to:4,operation:(embeds,state,positions)=>this.#positioned(mediaNarrow,ops.contiguous(embeds),state,"window",positions)},
|
|
933
|
+
{from:5,to:8,operation:(embeds,state,positions)=>this.#positionedBounded(mediaFiveToEight,ops.contiguous(embeds),state,positions)},
|
|
934
|
+
{from:9,to:15,operation:(embeds,state,positions)=>this.#positionedBounded(mediaWide,ops.contiguous(embeds),state,positions)},
|
|
935
|
+
{from:16,to:32,operation:(embeds,state,positions)=>this.#positionedBounded(mediaTiled,ops.contiguous(embeds),state,positions)},
|
|
936
|
+
{from:33,to:Infinity,operation:(embeds,state,positions)=>this.#positionedBounded(mediaExpanded,ops.contiguous(embeds),state,positions)},
|
|
937
|
+
];
|
|
938
|
+
this.mediaPrompt=[
|
|
939
|
+
{from:1,to:4,operation:(ids,state,positions)=>this.#positioned(mediaNarrow,this.embed.encode(ids),state,"window",positions)},
|
|
940
|
+
{from:5,to:8,operation:(ids,state,positions)=>this.#positionedBounded(mediaFiveToEight,this.embed.encode(ids),state,positions)},
|
|
941
|
+
{from:9,to:15,operation:(ids,state,positions)=>this.#positionedBounded(mediaWide,this.embed.encode(ids),state,positions)},
|
|
942
|
+
{from:16,to:32,operation:(ids,state,positions)=>this.#positionedBounded(mediaTiled,this.embed.encode(ids),state,positions)},
|
|
943
|
+
{from:33,to:Infinity,operation:(ids,state,positions)=>this.#positionedBounded(mediaExpanded,this.embed.encode(ids),state,positions)},
|
|
944
|
+
];
|
|
945
|
+
const mediaSmallDecode:MlxPositionedForward=(ids,state,positions)=>this.#positioned(mediaNarrow,this.embed.encode(ids),state,"decode",positions);
|
|
946
|
+
const mediaLargeDecode:MlxPositionedForward=(ids,state,positions)=>this.#positioned(mediaFiveToEight,this.embed.encode(ids),state,"decode",positions);
|
|
947
|
+
const mediaWideDecode:MlxPositionedForward=(ids,state,positions)=>this.#positioned(mediaWide,this.embed.encode(ids),state,"decode",positions);
|
|
948
|
+
const mediaTiledDecode:MlxPositionedForward=(ids,state,positions)=>this.#positioned(mediaTiled,this.embed.encode(ids),state,"decode",positions);
|
|
949
|
+
const mediaExpandedDecode:MlxPositionedForward=(ids,state,positions)=>this.#positioned(mediaExpanded,this.embed.encode(ids),state,"decode",positions);
|
|
950
|
+
this.mediaDecode=Object.freeze({
|
|
951
|
+
1:mediaSmallDecode,2:mediaSmallDecode,3:mediaSmallDecode,4:mediaSmallDecode,5:mediaLargeDecode,6:mediaLargeDecode,7:mediaLargeDecode,8:mediaLargeDecode,
|
|
952
|
+
9:mediaWideDecode,10:mediaWideDecode,11:mediaWideDecode,12:mediaWideDecode,13:mediaWideDecode,14:mediaWideDecode,15:mediaWideDecode,
|
|
953
|
+
16:mediaTiledDecode,17:mediaTiledDecode,18:mediaTiledDecode,19:mediaTiledDecode,20:mediaTiledDecode,21:mediaTiledDecode,22:mediaTiledDecode,23:mediaTiledDecode,
|
|
954
|
+
24:mediaTiledDecode,25:mediaTiledDecode,26:mediaTiledDecode,27:mediaTiledDecode,28:mediaTiledDecode,29:mediaTiledDecode,30:mediaTiledDecode,31:mediaTiledDecode,32:mediaTiledDecode,
|
|
955
|
+
|
|
956
|
+
default: mediaExpandedDecode,
|
|
957
|
+
});
|
|
958
|
+
const separatedWindow=[
|
|
959
|
+
{from:1,to:1,operation:(group:Parameters<MlxSeparateMixed["window"][number]["operation"]>[0])=>oneWindow(group.ids,group.state,group.capture)},
|
|
960
|
+
{from:2,to:3,operation:(group:Parameters<MlxSeparateMixed["window"][number]["operation"]>[0])=>rowsWindow(group.ids,group.state,group.capture)},
|
|
961
|
+
{from:4,to:4,operation:(group:Parameters<MlxSeparateMixed["window"][number]["operation"]>[0])=>fourWindow(group.ids,group.state,group.capture)},
|
|
962
|
+
{from:5,to:8,operation:(group:Parameters<MlxSeparateMixed["window"][number]["operation"]>[0])=>mmaWindow(group.ids,group.state,group.capture)},
|
|
963
|
+
{from:9,to:15,operation:(group:Parameters<MlxSeparateMixed["window"][number]["operation"]>[0])=>wideWindow(group.ids,group.state,group.capture)},
|
|
964
|
+
{from:16,to:32,operation:(group:Parameters<MlxSeparateMixed["window"][number]["operation"]>[0])=>tiledWindow(group.ids,group.state,group.capture)},
|
|
965
|
+
{from:33,to:Infinity,operation:(group:Parameters<MlxSeparateMixed["window"][number]["operation"]>[0])=>expandedWindow(group.ids,group.state,group.capture)},
|
|
966
|
+
];
|
|
967
|
+
this.embeddingPrompt=[
|
|
968
|
+
{from:1,to:1,operation:(hidden,cache)=>this.#embeddingSubmitted(this.#one,hidden,cache)},
|
|
969
|
+
{from:2,to:3,operation:(hidden,cache)=>this.#embeddingSubmitted(this.#rows,hidden,cache)},
|
|
970
|
+
{from:4,to:4,operation:(hidden,cache)=>this.#embeddingSubmitted(this.#four,hidden,cache)},
|
|
971
|
+
{from:5,to:8,operation:(hidden,cache)=>this.#embeddingSubmitted(this.#mma,hidden,cache)},
|
|
972
|
+
{from:9,to:15,operation:(hidden,cache)=>this.#embeddingBounded(this.#wide,hidden,cache)},
|
|
973
|
+
{from:16,to:32,operation:(hidden,cache)=>this.#embeddingBounded(this.#tiled,hidden,cache)},
|
|
974
|
+
{from:33,to:Infinity,operation:(hidden,cache)=>this.#embeddingBounded(this.#expanded,hidden,cache)},
|
|
975
|
+
];
|
|
976
|
+
this.mixedSeparate={decode:[
|
|
977
|
+
{from:1,to:1,operation:group=>oneDecode(group.ids,group.state,group.capture)},
|
|
978
|
+
{from:2,to:3,operation:group=>rowsDecode(group.ids,group.state,group.capture)},
|
|
979
|
+
{from:4,to:4,operation:group=>fourDecode(group.ids,group.state,group.capture)},
|
|
980
|
+
{from:5,to:8,operation:group=>mmaDecode(group.ids,group.state,group.capture)},
|
|
981
|
+
{from:9,to:15,operation:group=>wideDecode(group.ids,group.state,group.capture)},
|
|
982
|
+
{from:16,to:32,operation:group=>tiledDecode(group.ids,group.state,group.capture)},
|
|
983
|
+
{from:33,to:Infinity,operation:group=>expandedDecode(group.ids,group.state,group.capture)},
|
|
984
|
+
],window:separatedWindow,verify:separatedWindow};
|
|
985
|
+
console.error(`[graph] ${this.graph}`);
|
|
986
|
+
}
|
|
987
|
+
|
|
988
|
+
dispose():void { disposeResources(this.#owned.splice(0)); }
|
|
989
|
+
loraTargets():Map<string,QwenProjectionTarget> { return new Map(this.#targets); }
|
|
990
|
+
bindMediaInput(input:Vision):MlxPromptInput { return bindQwenMediaInput(this,input.embeddings,input.mrope!); }
|
|
991
|
+
mediaEncoders(modelDir:string):Promise<MediaEncoders> { return qwenMediaEncoders(this,modelDir); }
|
|
992
|
+
draftTarget(_state:Cache[]):TargetView { return qwenDraftTarget(this,this.#residualBasis); }
|
|
993
|
+
logitsFromHidden(hidden:MlxArray):MlxArray { return this.#head.forward(hidden); }
|
|
994
|
+
|
|
995
|
+
#submitted(layers:Layers,ids:MlxArray,state:Cache[],phase:ReadPhase,capture:MlxLayerCapture|null):MlxArray {
|
|
996
|
+
using scope=new SubmittedGraphForwardScope(this.embed.encode(ids),capture,64,4);
|
|
997
|
+
return this.#run(layers,state,phase,scope);
|
|
998
|
+
}
|
|
999
|
+
#embeddingSubmitted(layers:Layers,hidden:MlxArray,state:Cache[]):MlxArray {
|
|
1000
|
+
using scope=new SubmittedGraphForwardScope(ops.contiguous(hidden),null,64,4);
|
|
1001
|
+
return this.#run(layers,state,"window",scope);
|
|
1002
|
+
}
|
|
1003
|
+
#embeddingBounded(layers:Layers,hidden:MlxArray,state:Cache[]):MlxArray {
|
|
1004
|
+
using scope=new BoundedGraphForwardScope(ops.contiguous(hidden),null,state);
|
|
1005
|
+
return this.#run(layers,state,"window",scope);
|
|
1006
|
+
}
|
|
1007
|
+
#bounded(layers:Layers,ids:MlxArray,state:Cache[],phase:ReadPhase,capture:MlxLayerCapture|null):MlxArray {
|
|
1008
|
+
using scope=new BoundedGraphForwardScope(this.embed.encode(ids),capture,state);
|
|
1009
|
+
return this.#run(layers,state,phase,scope);
|
|
1010
|
+
}
|
|
1011
|
+
#lazy(layers:Layers,ids:MlxArray,state:Cache[],phase:ReadPhase,capture:MlxLayerCapture|null):MlxArray {
|
|
1012
|
+
using scope=new GraphForwardScope(this.embed.encode(ids),capture);
|
|
1013
|
+
return this.#run(layers,state,phase,scope);
|
|
1014
|
+
}
|
|
1015
|
+
#positioned(layers:Layers,hidden:MlxArray,state:Cache[],phase:ReadPhase,positions:MlxArray):MlxArray {
|
|
1016
|
+
using scope=new GraphForwardScope(hidden,null);
|
|
1017
|
+
return this.#run(layers,state,phase,scope,{posIds:positions,...this.#mrope});
|
|
1018
|
+
}
|
|
1019
|
+
#positionedBounded(layers:Layers,hidden:MlxArray,state:Cache[],positions:MlxArray):MlxArray {
|
|
1020
|
+
using scope=new BoundedGraphForwardScope(hidden,null,state);
|
|
1021
|
+
return this.#run(layers,state,"window",scope,{posIds:positions,...this.#mrope});
|
|
1022
|
+
}
|
|
1023
|
+
|
|
1024
|
+
/** Every callsite has its written DeltaNet or full-attention layer type. */
|
|
1025
|
+
#run(layers:Layers,state:Cache[],phase:ReadPhase,scope:GraphForwardScope,positions:MropeForwardState|null=null):MlxArray {
|
|
1026
|
+
scope.accept(0,layers[0].forward(scope.hidden,state[0]!,phase));
|
|
1027
|
+
scope.accept(1,layers[1].forward(scope.hidden,state[1]!,phase));
|
|
1028
|
+
scope.accept(2,layers[2].forward(scope.hidden,state[2]!,phase));
|
|
1029
|
+
scope.accept(3,layers[3].forward(scope.hidden,state[3]!,phase,positions));
|
|
1030
|
+
scope.accept(4,layers[4].forward(scope.hidden,state[4]!,phase));
|
|
1031
|
+
scope.accept(5,layers[5].forward(scope.hidden,state[5]!,phase));
|
|
1032
|
+
scope.accept(6,layers[6].forward(scope.hidden,state[6]!,phase));
|
|
1033
|
+
scope.accept(7,layers[7].forward(scope.hidden,state[7]!,phase,positions));
|
|
1034
|
+
scope.accept(8,layers[8].forward(scope.hidden,state[8]!,phase));
|
|
1035
|
+
scope.accept(9,layers[9].forward(scope.hidden,state[9]!,phase));
|
|
1036
|
+
scope.accept(10,layers[10].forward(scope.hidden,state[10]!,phase));
|
|
1037
|
+
scope.accept(11,layers[11].forward(scope.hidden,state[11]!,phase,positions));
|
|
1038
|
+
scope.accept(12,layers[12].forward(scope.hidden,state[12]!,phase));
|
|
1039
|
+
scope.accept(13,layers[13].forward(scope.hidden,state[13]!,phase));
|
|
1040
|
+
scope.accept(14,layers[14].forward(scope.hidden,state[14]!,phase));
|
|
1041
|
+
scope.accept(15,layers[15].forward(scope.hidden,state[15]!,phase,positions));
|
|
1042
|
+
scope.accept(16,layers[16].forward(scope.hidden,state[16]!,phase));
|
|
1043
|
+
scope.accept(17,layers[17].forward(scope.hidden,state[17]!,phase));
|
|
1044
|
+
scope.accept(18,layers[18].forward(scope.hidden,state[18]!,phase));
|
|
1045
|
+
scope.accept(19,layers[19].forward(scope.hidden,state[19]!,phase,positions));
|
|
1046
|
+
scope.accept(20,layers[20].forward(scope.hidden,state[20]!,phase));
|
|
1047
|
+
scope.accept(21,layers[21].forward(scope.hidden,state[21]!,phase));
|
|
1048
|
+
scope.accept(22,layers[22].forward(scope.hidden,state[22]!,phase));
|
|
1049
|
+
scope.accept(23,layers[23].forward(scope.hidden,state[23]!,phase,positions));
|
|
1050
|
+
scope.accept(24,layers[24].forward(scope.hidden,state[24]!,phase));
|
|
1051
|
+
scope.accept(25,layers[25].forward(scope.hidden,state[25]!,phase));
|
|
1052
|
+
scope.accept(26,layers[26].forward(scope.hidden,state[26]!,phase));
|
|
1053
|
+
scope.accept(27,layers[27].forward(scope.hidden,state[27]!,phase,positions));
|
|
1054
|
+
scope.accept(28,layers[28].forward(scope.hidden,state[28]!,phase));
|
|
1055
|
+
scope.accept(29,layers[29].forward(scope.hidden,state[29]!,phase));
|
|
1056
|
+
scope.accept(30,layers[30].forward(scope.hidden,state[30]!,phase));
|
|
1057
|
+
scope.accept(31,layers[31].forward(scope.hidden,state[31]!,phase,positions));
|
|
1058
|
+
scope.accept(32,layers[32].forward(scope.hidden,state[32]!,phase));
|
|
1059
|
+
scope.accept(33,layers[33].forward(scope.hidden,state[33]!,phase));
|
|
1060
|
+
scope.accept(34,layers[34].forward(scope.hidden,state[34]!,phase));
|
|
1061
|
+
scope.accept(35,layers[35].forward(scope.hidden,state[35]!,phase,positions));
|
|
1062
|
+
scope.accept(36,layers[36].forward(scope.hidden,state[36]!,phase));
|
|
1063
|
+
scope.accept(37,layers[37].forward(scope.hidden,state[37]!,phase));
|
|
1064
|
+
scope.accept(38,layers[38].forward(scope.hidden,state[38]!,phase));
|
|
1065
|
+
scope.accept(39,layers[39].forward(scope.hidden,state[39]!,phase,positions));
|
|
1066
|
+
scope.accept(40,layers[40].forward(scope.hidden,state[40]!,phase));
|
|
1067
|
+
scope.accept(41,layers[41].forward(scope.hidden,state[41]!,phase));
|
|
1068
|
+
scope.accept(42,layers[42].forward(scope.hidden,state[42]!,phase));
|
|
1069
|
+
scope.accept(43,layers[43].forward(scope.hidden,state[43]!,phase,positions));
|
|
1070
|
+
scope.accept(44,layers[44].forward(scope.hidden,state[44]!,phase));
|
|
1071
|
+
scope.accept(45,layers[45].forward(scope.hidden,state[45]!,phase));
|
|
1072
|
+
scope.accept(46,layers[46].forward(scope.hidden,state[46]!,phase));
|
|
1073
|
+
scope.accept(47,layers[47].forward(scope.hidden,state[47]!,phase,positions));
|
|
1074
|
+
scope.accept(48,layers[48].forward(scope.hidden,state[48]!,phase));
|
|
1075
|
+
scope.accept(49,layers[49].forward(scope.hidden,state[49]!,phase));
|
|
1076
|
+
scope.accept(50,layers[50].forward(scope.hidden,state[50]!,phase));
|
|
1077
|
+
scope.accept(51,layers[51].forward(scope.hidden,state[51]!,phase,positions));
|
|
1078
|
+
scope.accept(52,layers[52].forward(scope.hidden,state[52]!,phase));
|
|
1079
|
+
scope.accept(53,layers[53].forward(scope.hidden,state[53]!,phase));
|
|
1080
|
+
scope.accept(54,layers[54].forward(scope.hidden,state[54]!,phase));
|
|
1081
|
+
scope.accept(55,layers[55].forward(scope.hidden,state[55]!,phase,positions));
|
|
1082
|
+
scope.accept(56,layers[56].forward(scope.hidden,state[56]!,phase));
|
|
1083
|
+
scope.accept(57,layers[57].forward(scope.hidden,state[57]!,phase));
|
|
1084
|
+
scope.accept(58,layers[58].forward(scope.hidden,state[58]!,phase));
|
|
1085
|
+
scope.accept(59,layers[59].forward(scope.hidden,state[59]!,phase,positions));
|
|
1086
|
+
scope.accept(60,layers[60].forward(scope.hidden,state[60]!,phase));
|
|
1087
|
+
scope.accept(61,layers[61].forward(scope.hidden,state[61]!,phase));
|
|
1088
|
+
scope.accept(62,layers[62].forward(scope.hidden,state[62]!,phase));
|
|
1089
|
+
scope.accept(63,layers[63].forward(scope.hidden,state[63]!,phase,positions));
|
|
1090
|
+
return scope.finish(64,this.finalNorm.forward(scope.hidden));
|
|
1091
|
+
}
|
|
1092
|
+
}
|