doppler-gpu 0.4.7 → 0.4.8
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/CHANGELOG.md +64 -0
- package/README.md +93 -144
- package/models/catalog.json +1575 -0
- package/package.json +37 -12
- package/src/cli/doppler-cli.js +28 -7
- package/src/client/doppler-registry.json +2 -2
- package/src/client/runtime/index.js +25 -2
- package/src/client/runtime/model-session.d.ts +31 -1
- package/src/client/runtime/model-session.js +31 -0
- package/src/client/runtime/model-source.d.ts +19 -2
- package/src/client/runtime/model-source.js +45 -5
- package/src/config/conversion/gemma3/translategemma-4b-1b-enes-q4k-ehf16-af32.json +21 -5
- package/src/config/conversion/gemma3/translategemma-4b-it-q4k-ehf16-af32.json +6 -1
- package/src/config/conversion/gemma4/gemma-4-e2b-it-q4k-ehf16-af16-int4ple.json +1 -1
- package/src/config/conversion/qwen3/qwen-3-reranker-0-6b-q4k-ehf16-af32.json +9 -1
- package/src/config/kernel-path-loader.js +1 -0
- package/src/config/kernels/kernel-ref-digests.js +26 -17
- package/src/config/kernels/registry.json +448 -27
- package/src/config/merge.js +11 -0
- package/src/config/param-validator.js +25 -0
- package/src/config/runtime/default.json +3 -1
- package/src/config/runtime/profiles/qwen-3-5-0-8b-metal-attn128-candidate.json +20 -0
- package/src/config/runtime/profiles/qwen-3-5-0-8b-metal-maxbatch64-candidate.json +52 -0
- package/src/config/runtime/profiles/qwen-3-5-0-8b-metal-normed-ffn-candidate.json +18 -0
- package/src/config/runtime/profiles/qwen-3-5-0-8b-metal-overlap-ringtokens1-candidate.json +52 -0
- package/src/config/runtime/profiles/qwen-3-5-0-8b-metal-parity.json +85 -0
- package/src/config/runtime/profiles/qwen-3-5-0-8b-metal-sequential-candidate.json +52 -0
- package/src/config/runtime/profiles/qwen-3-5-0-8b-metal-throughput.json +85 -0
- package/src/config/runtime/profiles/qwen-3-5-2b-f16-primary-throughput.json +23 -0
- package/src/config/runtime/profiles/qwen-3-5-2b-metal-parity.json +38 -0
- package/src/config/runtime/profiles/qwen-3-5-2b-metal-throughput.json +135 -0
- package/src/config/runtime/profiles/qwen-3-embedding-0-6b-metal-stability.json +37 -0
- package/src/config/runtime/profiles/qwen-3-embedding-0-6b-metal-throughput.json +21 -0
- package/src/config/runtime/profiles/qwen-3-reranker-0-6b-metal-stability.json +41 -0
- package/src/config/runtime/profiles/qwen-3-reranker-0-6b-metal-throughput.json +41 -0
- package/src/config/runtime/profiles/rerank-stability.json +43 -0
- package/src/config/runtime/profiles/rerank-throughput.json +40 -0
- package/src/config/runtime/profiles/translategemma-1b-nativekd2-q4k-rdna3-throughput-probe.json +22 -0
- package/src/config/schema/doppler.schema.js +40 -5
- package/src/config/schema/execution-v1.schema.d.ts +41 -0
- package/src/config/schema/execution-v1.schema.js +10 -0
- package/src/config/schema/index.d.ts +3 -1
- package/src/config/schema/speculation-self.schema.d.ts +16 -0
- package/src/config/transforms/capability-transform-resolver.js +7 -1
- package/src/config/transforms/execution-graph-transforms.js +2 -2
- package/src/debug/config.d.ts +4 -0
- package/src/experimental/orchestration/index.d.ts +10 -1
- package/src/experimental/training/suite.js +1 -1
- package/src/formats/rdrr/classification.js +3 -0
- package/src/formats/rdrr/integrity.d.ts +0 -2
- package/src/formats/rdrr/integrity.js +0 -2
- package/src/formats/rdrr/types.d.ts +1 -1
- package/src/gpu/kernels/attention.d.ts +2 -0
- package/src/gpu/kernels/attention.js +75 -16
- package/src/gpu/kernels/attention_decode_online_head256_f16kv_output_gate.wgsl +268 -0
- package/src/gpu/kernels/attention_head128_f16kv.wgsl +269 -0
- package/src/gpu/kernels/fused_ffn.d.ts +32 -0
- package/src/gpu/kernels/fused_ffn.js +540 -19
- package/src/gpu/kernels/fused_ffn.wgsl +2 -1
- package/src/gpu/kernels/fused_ffn_f16.wgsl +2 -1
- package/src/gpu/kernels/fused_ffn_f16w_f32a_batched.wgsl +135 -0
- package/src/gpu/kernels/fused_ffn_q4k.wgsl +5 -3
- package/src/gpu/kernels/fused_ffn_q4k_f16.wgsl +2 -1
- package/src/gpu/kernels/fused_ffn_q4k_metal_simd16.wgsl +183 -0
- package/src/gpu/kernels/fused_gate_up_gelu_gated_f16.wgsl +2 -1
- package/src/gpu/kernels/fused_matmul_q4.wgsl +15 -6
- package/src/gpu/kernels/fused_matmul_q4_metal_simd16.wgsl +144 -0
- package/src/gpu/kernels/index.d.ts +21 -0
- package/src/gpu/kernels/index.js +13 -0
- package/src/gpu/kernels/linear-attention-core.js +413 -15
- package/src/gpu/kernels/lm-head-argmax.d.ts +10 -1
- package/src/gpu/kernels/lm-head-argmax.js +153 -57
- package/src/gpu/kernels/lm-head-select-logits.d.ts +23 -0
- package/src/gpu/kernels/lm-head-select-logits.js +186 -0
- package/src/gpu/kernels/lm_head_argmax_q4k.wgsl +289 -0
- package/src/gpu/kernels/lm_head_select_logits.wgsl +77 -0
- package/src/gpu/kernels/matmul-selection.js +5 -1
- package/src/gpu/kernels/residual_rmsnorm_pair.wgsl +127 -0
- package/src/gpu/kernels/rmsnorm-pair.d.ts +33 -0
- package/src/gpu/kernels/rmsnorm-pair.js +113 -0
- package/src/gpu/kernels/rmsnorm-stats.d.ts +29 -0
- package/src/gpu/kernels/rmsnorm-stats.js +250 -0
- package/src/gpu/kernels/silu.wgsl +7 -1
- package/src/gpu/kernels/silu_f16.wgsl +7 -1
- package/src/gpu/kernels/split-qkv-rmsnorm-rope-qk.js +15 -10
- package/src/gpu/kernels/split_qkv_rmsnorm_rope_qk.wgsl +37 -7
- package/src/gpu/kernels/split_qkv_rmsnorm_rope_qk_f16kv_cache.wgsl +37 -7
- package/src/gpu/tensor.d.ts +1 -0
- package/src/index.d.ts +1 -1
- package/src/index.js +1 -1
- package/src/inference/browser-harness-suite-helpers.js +1 -1
- package/src/inference/browser-harness-text-helpers.js +251 -20
- package/src/inference/browser-harness.js +221 -3
- package/src/inference/kv-cache/base.js +42 -1
- package/src/inference/pipelines/structured/json-grammar-mask.d.ts +18 -0
- package/src/inference/pipelines/text/attention/dispatch-params.js +8 -3
- package/src/inference/pipelines/text/attention/output-gate-fusion.js +37 -0
- package/src/inference/pipelines/text/attention/projections.js +174 -4
- package/src/inference/pipelines/text/attention/record.js +44 -21
- package/src/inference/pipelines/text/attention/run.js +18 -3
- package/src/inference/pipelines/text/config.d.ts +18 -0
- package/src/inference/pipelines/text/config.js +100 -0
- package/src/inference/pipelines/text/execution-plan.d.ts +2 -0
- package/src/inference/pipelines/text/execution-plan.js +37 -0
- package/src/inference/pipelines/text/execution-v1.d.ts +6 -0
- package/src/inference/pipelines/text/execution-v1.js +122 -8
- package/src/inference/pipelines/text/ffn/dense.js +239 -21
- package/src/inference/pipelines/text/ffn/standard.js +121 -38
- package/src/inference/pipelines/text/generator-decode-policy.js +1 -0
- package/src/inference/pipelines/text/generator-helpers.js +2 -0
- package/src/inference/pipelines/text/generator-runtime.js +1 -0
- package/src/inference/pipelines/text/generator-steps.js +2 -2
- package/src/inference/pipelines/text/generator.d.ts +18 -1
- package/src/inference/pipelines/text/generator.js +297 -19
- package/src/inference/pipelines/text/init-qkv-fusion.js +155 -62
- package/src/inference/pipelines/text/init.d.ts +22 -3
- package/src/inference/pipelines/text/init.js +2 -2
- package/src/inference/pipelines/text/layer.js +87 -8
- package/src/inference/pipelines/text/linear-attention-ab-fusion.js +208 -0
- package/src/inference/pipelines/text/linear-attention.d.ts +1 -0
- package/src/inference/pipelines/text/linear-attention.js +173 -16
- package/src/inference/pipelines/text/logits/gpu.js +2 -2
- package/src/inference/pipelines/text/logits/index.js +62 -1
- package/src/inference/pipelines/text/model-load.js +6 -0
- package/src/inference/pipelines/text/ops.js +22 -0
- package/src/inference/pipelines/text/per-layer-inputs.js +80 -11
- package/src/inference/pipelines/text/probes.js +9 -4
- package/src/inference/pipelines/text/state.d.ts +4 -4
- package/src/inference/pipelines/text/state.js +4 -0
- package/src/inference/pipelines/text/types.d.ts +27 -2
- package/src/inference/pipelines/text.d.ts +21 -1
- package/src/inference/pipelines/text.js +40 -14
- package/src/inference/tokenizers/bundled.js +45 -23
- package/src/inference/tokenizers/types.d.ts +7 -4
- package/src/loader/doppler-loader.js +25 -2
- package/src/loader/index.d.ts +3 -1
- package/src/loader/index.js +2 -0
- package/src/loader/loader-types.d.ts +3 -1
- package/src/loader/manifest-config.d.ts +18 -0
- package/src/loader/manifest-config.js +11 -3
- package/src/loader/shard-cache.d.ts +2 -0
- package/src/loader/shard-cache.js +29 -1
- package/src/models/gemma4.js +8 -0
- package/src/rules/inference/capability-transforms.rules.json +145 -20
- package/src/rules/inference/ffn.rules.json +33 -0
- package/src/rules/kernels/fused-ffn.rules.json +1 -1
- package/src/rules/kernels/lm-head-argmax.rules.json +10 -0
- package/src/rules/kernels/matmul.rules.json +1 -0
- package/src/rules/kernels/silu.rules.json +1 -0
- package/src/rules/rule-registry.js +2 -0
- package/src/storage/artifact-storage-context.js +36 -0
- package/src/tooling/node-browser-command-runner.js +42 -7
- package/src/tooling/opfs-cache.d.ts +21 -1
- package/src/tooling/opfs-cache.js +168 -32
- package/src/tooling-exports/storage.d.ts +1 -1
- package/src/tooling-exports/storage.js +1 -1
- package/src/types/gpu.d.ts +25 -0
- package/src/version.js +1 -1
- package/tests/kernels/browser/kernel-suite.js +1593 -0
- package/tests/kernels/browser/test-page.js +2629 -0
- package/tests/kernels/harness/benchmark.js +225 -0
- package/tests/kernels/harness/buffer-utils.js +88 -0
- package/tests/kernels/harness/tolerance.js +174 -0
- package/tests/kernels/reference/adam.js +15 -0
- package/tests/kernels/reference/attention.js +158 -0
- package/tests/kernels/reference/backward.js +96 -0
- package/tests/kernels/reference/bias-add.js +20 -0
- package/tests/kernels/reference/conv2d.js +58 -0
- package/tests/kernels/reference/cross-entropy.js +26 -0
- package/tests/kernels/reference/dequant.js +312 -0
- package/tests/kernels/reference/embed.js +15 -0
- package/tests/kernels/reference/gather.js +61 -0
- package/tests/kernels/reference/gelu.js +37 -0
- package/tests/kernels/reference/groupnorm.js +36 -0
- package/tests/kernels/reference/index.js +89 -0
- package/tests/kernels/reference/layernorm.js +22 -0
- package/tests/kernels/reference/matmul.js +57 -0
- package/tests/kernels/reference/modulate.js +24 -0
- package/tests/kernels/reference/moe-gather.js +83 -0
- package/tests/kernels/reference/pixel-shuffle.js +20 -0
- package/tests/kernels/reference/residual.js +55 -0
- package/tests/kernels/reference/rmsnorm.js +52 -0
- package/tests/kernels/reference/rope.js +79 -0
- package/tests/kernels/reference/sample.js +97 -0
- package/tests/kernels/reference/scatter-add.js +58 -0
- package/tests/kernels/reference/silu.js +52 -0
- package/tests/kernels/reference/softmax.js +86 -0
- package/tests/kernels/reference/split-qkv.js +58 -0
- package/tests/kernels/reference/topk.js +88 -0
- package/tests/kernels/reference/upsample2d.js +37 -0
- package/BRANDING.md +0 -14
- package/SECURITY.md +0 -19
- package/assets/doppler-runtime-map.svg +0 -82
- package/src/bootstrap.js +0 -30
- package/src/cli/doppler-cli.d.ts +0 -46
- package/src/cli/doppler-quickstart.d.ts +0 -45
- package/src/cli/doppler-serve.d.ts +0 -210
- package/src/cli/runtime-profiles.d.ts +0 -40
- package/src/client/doppler-api.browser.d.ts +0 -1
- package/src/client/doppler-api.browser.js +0 -25
- package/src/client/doppler-registry.d.ts +0 -54
- package/src/client/runtime/node-quickstart-cache.d.ts +0 -22
- package/src/client/runtime/source-runtime.d.ts +0 -26
- package/src/client/runtime/storage.d.ts +0 -5
- package/src/config/README.md +0 -61
- package/src/config/conversion/README.md +0 -62
- package/src/config/goal-completion-matrix.json +0 -552
- package/src/config/kernels/kernel-ref-digests.d.ts +0 -1
- package/src/config/kernels/moe/gpt-oss.paths.json +0 -49
- package/src/config/kernels/moe/mixtral.paths.json +0 -46
- package/src/config/kernels/registry.d.ts +0 -51
- package/src/config/merge-contract-check.d.ts +0 -16
- package/src/config/merge-contract-check.js +0 -319
- package/src/config/merge-helpers.d.ts +0 -60
- package/src/config/model-lanes/catalog-lane-resolver.d.ts +0 -66
- package/src/config/model-lanes/catalog-lane-resolver.js +0 -170
- package/src/config/param-categories.d.ts +0 -8
- package/src/config/param-validator.d.ts +0 -33
- package/src/config/platforms/loader.d.ts +0 -38
- package/src/config/quantization-contract-check.d.ts +0 -12
- package/src/config/quantization-contract-check.js +0 -91
- package/src/config/runtime/README.md +0 -28
- package/src/config/runtime-merge.d.ts +0 -5
- package/src/config/schema/claim-evidence-contract.schema.json +0 -93
- package/src/config/schema/command-surface-contract.schema.json +0 -81
- package/src/config/schema/goal-completion-matrix.schema.json +0 -223
- package/src/config/schema/model-artifact-contract.schema.json +0 -119
- package/src/config/schema/platform.schema.js +0 -1
- package/src/config/schema/policy-schema-registry.json +0 -31
- package/src/config/schema/subsystem-support-contract.schema.json +0 -48
- package/src/config/support-tiers/subsystems.json +0 -502
- package/src/config/supported-operations.d.ts +0 -14
- package/src/config/transforms/capability-transform-resolver.d.ts +0 -56
- package/src/config/transforms/execution-graph-transforms.d.ts +0 -306
- package/src/converter/conversion-plan.d.ts +0 -56
- package/src/converter/index.d.ts +0 -98
- package/src/converter/index.js +0 -59
- package/src/converter/parsers/diffusion.d.ts +0 -50
- package/src/converter/parsers/gguf.d.ts +0 -31
- package/src/converter/parsers/index.d.ts +0 -21
- package/src/converter/parsers/index.js +0 -12
- package/src/converter/parsers/transformer.d.ts +0 -31
- package/src/converter/quantization-info.d.ts +0 -38
- package/src/converter/rope-config.d.ts +0 -16
- package/src/converter/tokenizer-utils.d.ts +0 -12
- package/src/debug/reference/README.md +0 -114
- package/src/debug/reference/hf_attn_debug.py +0 -114
- package/src/debug/reference/hf_embed_check.py +0 -89
- package/src/debug/reference/hf_layer_out.py +0 -100
- package/src/debug/reference/hf_qwen35_linear_attn_debug.py +0 -268
- package/src/debug/reference/hf_rope_check.py +0 -116
- package/src/debug/reference/hf_weights.py +0 -75
- package/src/errors/doppler-error.d.ts +0 -20
- package/src/experimental/bridge/extension/background.d.ts +0 -14
- package/src/experimental/bridge/extension/background.js +0 -168
- package/src/experimental/bridge/extension/manifest.json +0 -34
- package/src/experimental/bridge/extension-client.d.ts +0 -114
- package/src/experimental/bridge/index.d.ts +0 -69
- package/src/experimental/bridge/node-bridge.js +0 -24
- package/src/experimental/bridge/protocol.d.ts +0 -96
- package/src/experimental/browser/gguf-importer.d.ts +0 -136
- package/src/experimental/browser/gguf-importer.js +0 -532
- package/src/experimental/browser/gguf-parser-browser.d.ts +0 -14
- package/src/experimental/browser/quantization.d.ts +0 -74
- package/src/experimental/browser/safetensors-parser-browser.d.ts +0 -193
- package/src/experimental/browser/shard-io-browser.d.ts +0 -57
- package/src/experimental/browser/shard-io-browser.js +0 -89
- package/src/experimental/browser/tensor-source-download.d.ts +0 -27
- package/src/experimental/diffusion/config/expansion-lanes.json +0 -29
- package/src/experimental/distribution/p2p-transport-contract.d.ts +0 -57
- package/src/experimental/distribution/p2p-transport.js +0 -75
- package/src/experimental/hotswap/intent-bundle.d.ts +0 -37
- package/src/experimental/hotswap/manifest.d.ts +0 -42
- package/src/experimental/hotswap/model-swap.js +0 -42
- package/src/experimental/hotswap/runtime.d.ts +0 -31
- package/src/experimental/training/README.md +0 -164
- package/src/experimental/training/artifacts.d.ts +0 -160
- package/src/experimental/training/attention-backward.d.ts +0 -30
- package/src/experimental/training/attention-forward.d.ts +0 -22
- package/src/experimental/training/attention-forward.js +0 -82
- package/src/experimental/training/autograd.d.ts +0 -51
- package/src/experimental/training/backward-kernels.js +0 -39
- package/src/experimental/training/checkpoint-watch.d.ts +0 -8
- package/src/experimental/training/checkpoint.d.ts +0 -36
- package/src/experimental/training/clip.d.ts +0 -9
- package/src/experimental/training/dataloader.d.ts +0 -8
- package/src/experimental/training/datasets/gepa-frontier.d.ts +0 -41
- package/src/experimental/training/datasets/gepa-frontier.js +0 -220
- package/src/experimental/training/datasets/index.d.ts +0 -24
- package/src/experimental/training/datasets/index.js +0 -13
- package/src/experimental/training/datasets/jsonl.d.ts +0 -11
- package/src/experimental/training/datasets/reploid.d.ts +0 -3
- package/src/experimental/training/datasets/reploid.js +0 -36
- package/src/experimental/training/datasets/teacher-traces.d.ts +0 -51
- package/src/experimental/training/datasets/teacher-traces.js +0 -279
- package/src/experimental/training/datasets/text-pairs.d.ts +0 -66
- package/src/experimental/training/datasets/token-batch.d.ts +0 -21
- package/src/experimental/training/datasets/token-batch.js +0 -52
- package/src/experimental/training/datasets/translation-pairs.d.ts +0 -34
- package/src/experimental/training/datasets/translation-pairs.js +0 -49
- package/src/experimental/training/distillation/artifacts.d.ts +0 -71
- package/src/experimental/training/distillation/checkpoint-watch.d.ts +0 -10
- package/src/experimental/training/distillation/dataset.d.ts +0 -59
- package/src/experimental/training/distillation/eval.d.ts +0 -34
- package/src/experimental/training/distillation/index.d.ts +0 -29
- package/src/experimental/training/distillation/runtime.d.ts +0 -20
- package/src/experimental/training/distillation/scoreboard.d.ts +0 -6
- package/src/experimental/training/distillation/stage-a.d.ts +0 -45
- package/src/experimental/training/distillation/stage-b.d.ts +0 -24
- package/src/experimental/training/distillation/student-fixture.d.ts +0 -27
- package/src/experimental/training/distillation/suite-data.d.ts +0 -45
- package/src/experimental/training/export.d.ts +0 -41
- package/src/experimental/training/index.d.ts +0 -74
- package/src/experimental/training/index.js +0 -63
- package/src/experimental/training/lora-pipeline.d.ts +0 -183
- package/src/experimental/training/lora.d.ts +0 -19
- package/src/experimental/training/loss-scaling.d.ts +0 -21
- package/src/experimental/training/loss.d.ts +0 -10
- package/src/experimental/training/objectives/base.d.ts +0 -58
- package/src/experimental/training/objectives/cross_entropy.d.ts +0 -18
- package/src/experimental/training/objectives/distill_kd.d.ts +0 -16
- package/src/experimental/training/objectives/distill_triplet.d.ts +0 -16
- package/src/experimental/training/objectives/index.d.ts +0 -12
- package/src/experimental/training/objectives/index.js +0 -6
- package/src/experimental/training/objectives/ul_stage1_joint.d.ts +0 -16
- package/src/experimental/training/objectives/ul_stage2_base.d.ts +0 -16
- package/src/experimental/training/operator-agent-eval.d.ts +0 -65
- package/src/experimental/training/operator-artifacts.d.ts +0 -62
- package/src/experimental/training/operator-command.d.ts +0 -5
- package/src/experimental/training/operator-eval.d.ts +0 -48
- package/src/experimental/training/operator-scoreboard.d.ts +0 -5
- package/src/experimental/training/optimizer.d.ts +0 -22
- package/src/experimental/training/runner.d.ts +0 -249
- package/src/experimental/training/suite.d.ts +0 -299
- package/src/experimental/training/tensor-factory.d.ts +0 -9
- package/src/experimental/training/trainer.d.ts +0 -89
- package/src/experimental/training/ul_dataset.d.ts +0 -47
- package/src/experimental/training/ul_schedule.d.ts +0 -6
- package/src/experimental/training/workload-packs/distill-glm52-doppler-agent-harness-gemma270m-f16-sft-tiny.json +0 -160
- package/src/experimental/training/workload-packs/distill-glm52-doppler-code-agent-gemma270m-f16-sft-tiny.json +0 -175
- package/src/experimental/training/workload-packs/distill-glm52-doppler-code-agent-sft-tiny.json +0 -175
- package/src/experimental/training/workload-packs/distill-glm52-doppler-js-json-gemma270m-f16-sft-tiny.json +0 -160
- package/src/experimental/training/workload-packs/distill-glm52-doppler-review-gemma270m-f16-sft-tiny.json +0 -162
- package/src/experimental/training/workload-packs/distill-glm52-doppler-wgsl-gemma270m-f16-sft-tiny.json +0 -160
- package/src/experimental/training/workload-packs/distill-translategemma-medium.json +0 -103
- package/src/experimental/training/workload-packs/distill-translategemma-tiny.json +0 -103
- package/src/experimental/training/workload-packs/lora-doppler-agent-harness-gemma270m-f16-tiny.json +0 -137
- package/src/experimental/training/workload-packs/lora-doppler-code-agent-gemma270m-f16-tiny.json +0 -152
- package/src/experimental/training/workload-packs/lora-doppler-code-agent-tiny.json +0 -152
- package/src/experimental/training/workload-packs/lora-doppler-js-json-gemma270m-f16-tiny.json +0 -137
- package/src/experimental/training/workload-packs/lora-doppler-review-gemma270m-f16-tiny.json +0 -139
- package/src/experimental/training/workload-packs/lora-doppler-wgsl-gemma270m-f16-tiny.json +0 -137
- package/src/experimental/training/workload-packs/lora-toy-medium.json +0 -91
- package/src/experimental/training/workload-packs/lora-toy-tiny.json +0 -91
- package/src/experimental/training/workload-packs/registry.json +0 -278
- package/src/experimental/training/workload-packs/ul-training-medium.json +0 -55
- package/src/experimental/training/workload-packs/ul-training-tiny.json +0 -55
- package/src/experimental/training/workloads.d.ts +0 -201
- package/src/formats/gguf/types.d.ts +0 -74
- package/src/formats/index.d.ts +0 -75
- package/src/formats/index.js +0 -19
- package/src/formats/litert/types.d.ts +0 -74
- package/src/formats/rdrr/distributed/parsing.d.ts +0 -6
- package/src/formats/rdrr/distributed/validation.d.ts +0 -23
- package/src/formats/rdrr/storage-descriptor.d.ts +0 -36
- package/src/formats/rdrr/tensor-config-validator.d.ts +0 -33
- package/src/formats/safetensors/types.d.ts +0 -67
- package/src/formats/tflite/types.d.ts +0 -61
- package/src/formats/tokenizer/index.d.ts +0 -5
- package/src/formats/tokenizer/index.js +0 -3
- package/src/formats/tokenizer/types.d.ts +0 -9
- package/src/gpu/kernel-runtime.d.ts +0 -13
- package/src/gpu/kernel-selection-cache.d.ts +0 -13
- package/src/gpu/kernel-selection-log.d.ts +0 -4
- package/src/gpu/kernel-tuner/benchmarks.d.ts +0 -144
- package/src/gpu/kernel-tuner/cache.d.ts +0 -55
- package/src/gpu/kernel-tuner/index.d.ts +0 -59
- package/src/gpu/kernel-tuner/tuner.d.ts +0 -82
- package/src/gpu/kernel-tuner/types.d.ts +0 -101
- package/src/gpu/kernel-tuner/types.js +0 -4
- package/src/gpu/kernel-tuner.d.ts +0 -33
- package/src/gpu/kernels/README.md +0 -166
- package/src/gpu/kernels/check-finiteness.d.ts +0 -15
- package/src/gpu/kernels/check-hot-vocab-stop.d.ts +0 -18
- package/src/gpu/kernels/check-stop.d.ts +0 -31
- package/src/gpu/kernels/codegen/patches/attention_decode_f16.from.attention_decode.diff +0 -213
- package/src/gpu/kernels/codegen/patches/attention_decode_f16kv.from.attention_decode.diff +0 -89
- package/src/gpu/kernels/codegen/patches/attention_f16.from.attention.diff +0 -62
- package/src/gpu/kernels/codegen/patches/attention_f16kv.from.attention.diff +0 -40
- package/src/gpu/kernels/codegen/patches/attention_small_f16.from.attention_small.diff +0 -226
- package/src/gpu/kernels/codegen/patches/attention_small_f16kv.from.attention_small.diff +0 -264
- package/src/gpu/kernels/codegen/patches/attention_streaming_f16.from.attention_streaming.diff +0 -71
- package/src/gpu/kernels/codegen/patches/attention_streaming_f16kv.from.attention_streaming.diff +0 -53
- package/src/gpu/kernels/codegen/patches/bias_add_f16.from.bias_add.diff +0 -54
- package/src/gpu/kernels/codegen/patches/dequant_mxfp4_expert_f16.from.dequant_mxfp4_expert.diff +0 -125
- package/src/gpu/kernels/codegen/patches/fused_ffn_f16.from.fused_ffn.diff +0 -428
- package/src/gpu/kernels/codegen/patches/fused_matmul_q4_batched_f16.from.fused_matmul_q4_batched.diff +0 -114
- package/src/gpu/kernels/codegen/patches/fused_matmul_rmsnorm_f16.from.fused_matmul_rmsnorm.diff +0 -477
- package/src/gpu/kernels/codegen/patches/fused_swiglu_f16.from.fused_swiglu.diff +0 -60
- package/src/gpu/kernels/codegen/patches/gelu_f16.from.gelu.diff +0 -41
- package/src/gpu/kernels/codegen/patches/layernorm_f16.from.layernorm.diff +0 -82
- package/src/gpu/kernels/codegen/patches/matmul_gemv_residual_f16.from.matmul_gemv_residual.diff +0 -87
- package/src/gpu/kernels/codegen/patches/moe_gather_f16.from.moe_gather.diff +0 -181
- package/src/gpu/kernels/codegen/patches/residual_f16.from.residual.diff +0 -65
- package/src/gpu/kernels/codegen/patches/rmsnorm_f16.from.rmsnorm.diff +0 -515
- package/src/gpu/kernels/codegen/patches/rope_f16.from.rope.diff +0 -292
- package/src/gpu/kernels/codegen/patches/sample_f16.from.sample.diff +0 -336
- package/src/gpu/kernels/codegen/patches/scatter_add_dynamic_f16.from.scatter_add_dynamic.diff +0 -65
- package/src/gpu/kernels/codegen/patches/silu_f16.from.silu.diff +0 -81
- package/src/gpu/kernels/codegen/patches/topk_f16.from.topk.diff +0 -231
- package/src/gpu/kernels/codegen/wgsl-patch-variants.js +0 -152
- package/src/gpu/kernels/codegen/wgsl-variants.js +0 -885
- package/src/gpu/kernels/constants.d.ts +0 -129
- package/src/gpu/kernels/dispatch.d.ts +0 -72
- package/src/gpu/kernels/gated-short-conv.d.ts +0 -63
- package/src/gpu/kernels/kernel-base.d.ts +0 -33
- package/src/gpu/kernels/linear-attention-core.d.ts +0 -42
- package/src/gpu/kernels/matmul-dispatch.d.ts +0 -49
- package/src/gpu/kernels/matmul-selection.d.ts +0 -87
- package/src/gpu/kernels/rep-penalty.d.ts +0 -21
- package/src/gpu/kernels/rule-matcher.d.ts +0 -30
- package/src/gpu/kernels/rule-registry.d.ts +0 -7
- package/src/gpu/kernels/types.js +0 -4
- package/src/gpu/perf-guards.d.ts +0 -13
- package/src/gpu/profiler.d.ts +0 -93
- package/src/gpu/readback-utils.d.ts +0 -16
- package/src/gpu/readback-utils.js +0 -49
- package/src/gpu/submit-probe.d.ts +0 -12
- package/src/index-browser.d.ts +0 -2
- package/src/index-browser.js +0 -2
- package/src/inference/README.md +0 -593
- package/src/inference/browser-harness-contract-helpers.d.ts +0 -5
- package/src/inference/browser-harness-diffusion-energy-suites.d.ts +0 -2
- package/src/inference/browser-harness-model-helpers.d.ts +0 -10
- package/src/inference/browser-harness-report-helpers.d.ts +0 -7
- package/src/inference/browser-harness-suite-helpers.d.ts +0 -40
- package/src/inference/browser-harness-text-helpers.d.ts +0 -152
- package/src/inference/model-lanes/execution-lane-resolver.d.ts +0 -61
- package/src/inference/model-lanes/execution-lane-resolver.js +0 -126
- package/src/inference/network-evolution.js +0 -79
- package/src/inference/pipelines/context.d.ts +0 -21
- package/src/inference/pipelines/diffusion/helpers.d.ts +0 -29
- package/src/inference/pipelines/diffusion/index.d.ts +0 -3
- package/src/inference/pipelines/diffusion/index.js +0 -3
- package/src/inference/pipelines/diffusion/scheduler.d.ts +0 -35
- package/src/inference/pipelines/diffusion/sd3-transformer.d.ts +0 -20
- package/src/inference/pipelines/diffusion/sd3-weights.d.ts +0 -21
- package/src/inference/pipelines/diffusion/text-encoder-gpu.d.ts +0 -82
- package/src/inference/pipelines/diffusion/text-encoder.d.ts +0 -29
- package/src/inference/pipelines/diffusion/types.js +0 -1
- package/src/inference/pipelines/diffusion/vae.d.ts +0 -20
- package/src/inference/pipelines/dream/energy-head-pipeline.d.ts +0 -29
- package/src/inference/pipelines/dream/energy-head-pipeline.js +0 -6
- package/src/inference/pipelines/dream/pipeline.d.ts +0 -17
- package/src/inference/pipelines/dream/pipeline.js +0 -8
- package/src/inference/pipelines/energy/index.d.ts +0 -1
- package/src/inference/pipelines/energy/index.js +0 -1
- package/src/inference/pipelines/energy/types.js +0 -1
- package/src/inference/pipelines/energy-head/index.d.ts +0 -6
- package/src/inference/pipelines/energy-head/index.js +0 -6
- package/src/inference/pipelines/factory.d.ts +0 -10
- package/src/inference/pipelines/index.d.ts +0 -22
- package/src/inference/pipelines/index.js +0 -19
- package/src/inference/pipelines/registry.d.ts +0 -15
- package/src/inference/pipelines/rng.d.ts +0 -2
- package/src/inference/pipelines/structured/tsv-grammar-mask.js +0 -166
- package/src/inference/pipelines/text/attention/attn-config.d.ts +0 -23
- package/src/inference/pipelines/text/attention/dispatch-params.d.ts +0 -173
- package/src/inference/pipelines/text/attention/index.d.ts +0 -24
- package/src/inference/pipelines/text/attention/output-projection.d.ts +0 -12
- package/src/inference/pipelines/text/attention/precision-contract.d.ts +0 -36
- package/src/inference/pipelines/text/attention/projections.d.ts +0 -208
- package/src/inference/pipelines/text/attention/quant-options.d.ts +0 -38
- package/src/inference/pipelines/text/attention/record.d.ts +0 -36
- package/src/inference/pipelines/text/attention/rmsnorm-fusion-gate.d.ts +0 -8
- package/src/inference/pipelines/text/attention/run.d.ts +0 -38
- package/src/inference/pipelines/text/attention/types.d.ts +0 -125
- package/src/inference/pipelines/text/attention.d.ts +0 -23
- package/src/inference/pipelines/text/attention.js +0 -12
- package/src/inference/pipelines/text/bdpa-steamroller.d.ts +0 -22
- package/src/inference/pipelines/text/buffer-types.js +0 -4
- package/src/inference/pipelines/text/debug-utils/config.d.ts +0 -144
- package/src/inference/pipelines/text/debug-utils/index.d.ts +0 -53
- package/src/inference/pipelines/text/debug-utils/logging.d.ts +0 -106
- package/src/inference/pipelines/text/debug-utils/tensor.d.ts +0 -119
- package/src/inference/pipelines/text/debug-utils/utils.d.ts +0 -77
- package/src/inference/pipelines/text/debug-utils.d.ts +0 -42
- package/src/inference/pipelines/text/drift-policy.d.ts +0 -57
- package/src/inference/pipelines/text/dtype-contract.d.ts +0 -53
- package/src/inference/pipelines/text/embed.d.ts +0 -69
- package/src/inference/pipelines/text/execution-runtime-builders.d.ts +0 -49
- package/src/inference/pipelines/text/failure-reduction.d.ts +0 -84
- package/src/inference/pipelines/text/failure-reduction.js +0 -180
- package/src/inference/pipelines/text/ffn/dense.d.ts +0 -99
- package/src/inference/pipelines/text/ffn/index.d.ts +0 -23
- package/src/inference/pipelines/text/ffn/moe.d.ts +0 -24
- package/src/inference/pipelines/text/ffn/sandwich.d.ts +0 -27
- package/src/inference/pipelines/text/ffn/standard.d.ts +0 -26
- package/src/inference/pipelines/text/ffn/types.d.ts +0 -30
- package/src/inference/pipelines/text/ffn.d.ts +0 -31
- package/src/inference/pipelines/text/ffn.js +0 -18
- package/src/inference/pipelines/text/finiteness-guard-status.d.ts +0 -11
- package/src/inference/pipelines/text/finiteness-policy.d.ts +0 -39
- package/src/inference/pipelines/text/generator-decode-policy.d.ts +0 -81
- package/src/inference/pipelines/text/generator-helpers.d.ts +0 -45
- package/src/inference/pipelines/text/generator-prefill-helpers.d.ts +0 -84
- package/src/inference/pipelines/text/generator-runtime.d.ts +0 -93
- package/src/inference/pipelines/text/generator-steps.d.ts +0 -186
- package/src/inference/pipelines/text/index.d.ts +0 -5
- package/src/inference/pipelines/text/index.js +0 -6
- package/src/inference/pipelines/text/kernel-trace.d.ts +0 -119
- package/src/inference/pipelines/text/layer.d.ts +0 -118
- package/src/inference/pipelines/text/load-errors.d.ts +0 -21
- package/src/inference/pipelines/text/logits/precision-policy.d.ts +0 -11
- package/src/inference/pipelines/text/logits/types.js +0 -4
- package/src/inference/pipelines/text/logits.d.ts +0 -27
- package/src/inference/pipelines/text/logits.js +0 -16
- package/src/inference/pipelines/text/lora-apply.d.ts +0 -28
- package/src/inference/pipelines/text/moe-cache.d.ts +0 -32
- package/src/inference/pipelines/text/moe-cpu-gptoss.d.ts +0 -9
- package/src/inference/pipelines/text/moe-cpu.d.ts +0 -13
- package/src/inference/pipelines/text/moe-gpu.d.ts +0 -26
- package/src/inference/pipelines/text/moe-helpers.d.ts +0 -12
- package/src/inference/pipelines/text/moe-shape-validator.d.ts +0 -70
- package/src/inference/pipelines/text/operator-events.d.ts +0 -125
- package/src/inference/pipelines/text/operator-identity.d.ts +0 -60
- package/src/inference/pipelines/text/ops.d.ts +0 -222
- package/src/inference/pipelines/text/per-layer-inputs.d.ts +0 -107
- package/src/inference/pipelines/text/probes.d.ts +0 -28
- package/src/inference/pipelines/text/sampling-config.d.ts +0 -48
- package/src/inference/pipelines/text/stage-names.d.ts +0 -88
- package/src/inference/pipelines/text/tsir-fixture-writer.d.ts +0 -57
- package/src/inference/pipelines/text/types.js +0 -4
- package/src/inference/tokenizers/base.d.ts +0 -44
- package/src/inference/tokenizers/behavior-flags.d.ts +0 -9
- package/src/inference/tokenizers/bpe.d.ts +0 -27
- package/src/inference/tokenizers/bundled.d.ts +0 -67
- package/src/inference/tokenizers/sentencepiece.d.ts +0 -30
- package/src/inference/tokenizers/types.js +0 -7
- package/src/loader/embedding-limit-preflight.d.ts +0 -31
- package/src/loader/per-layer-input-loader.d.ts +0 -30
- package/src/loader/tensors/functional-descriptor-loader.d.ts +0 -9
- package/src/loader/tensors/source-transform.d.ts +0 -65
- package/src/memory/address-table.d.ts +0 -96
- package/src/rules/execution-rules-contract-check.d.ts +0 -21
- package/src/rules/layer-pattern-contract-check.d.ts +0 -17
- package/src/rules/rule-registry.d.ts +0 -77
- package/src/storage/artifact-storage-context.d.ts +0 -63
- package/src/storage/backends/idb-store.d.ts +0 -52
- package/src/storage/backends/memory-store.d.ts +0 -36
- package/src/storage/backends/opfs-store.d.ts +0 -41
- package/src/storage/blake3.d.ts +0 -17
- package/src/storage/download-types.d.ts +0 -157
- package/src/storage/downloader.d.ts +0 -103
- package/src/storage/index.d.ts +0 -274
- package/src/storage/index.js +0 -201
- package/src/storage/preflight.d.ts +0 -144
- package/src/storage/preflight.js +0 -360
- package/src/storage/quickstart-downloader.d.ts +0 -157
- package/src/storage/quickstart-downloader.js +0 -289
- package/src/storage/source-artifact-store.d.ts +0 -59
- package/src/sw.js +0 -184
- package/src/tooling/command-api-constants.d.ts +0 -9
- package/src/tooling/command-api-family-normalizers.d.ts +0 -10
- package/src/tooling/command-api-helpers.d.ts +0 -23
- package/src/tooling/command-runner-shared.d.ts +0 -89
- package/src/tooling/conversion-config-materializer.d.ts +0 -14
- package/src/tooling/conversion-config-materializer.js +0 -96
- package/src/tooling/diagnose-runner.d.ts +0 -28
- package/src/tooling/hf-registry-utils.d.ts +0 -55
- package/src/tooling/kernel-path-builder/custom-runtime-facts.d.ts +0 -18
- package/src/tooling/kernel-path-builder/custom-runtime-facts.js +0 -72
- package/src/tooling/kernel-path-builder/index.d.ts +0 -65
- package/src/tooling/kernel-path-builder/index.js +0 -1361
- package/src/tooling/kernel-path-builder/runtime-overlay.d.ts +0 -9
- package/src/tooling/kernel-path-builder/runtime-overlay.js +0 -147
- package/src/tooling/lean-execution-contract-runner.d.ts +0 -33
- package/src/tooling/lean-execution-contract-runner.js +0 -158
- package/src/tooling/lean-execution-contract.d.ts +0 -16
- package/src/tooling/lean-execution-contract.js +0 -81
- package/src/tooling/litert-package-runtime.d.ts +0 -46
- package/src/tooling/node-convert-worker-pool.d.ts +0 -16
- package/src/tooling/node-convert-worker.d.ts +0 -1
- package/src/tooling/node-converter.d.ts +0 -1
- package/src/tooling/node-file-fetch.d.ts +0 -1
- package/src/tooling/node-source-runtime.d.ts +0 -24
- package/src/tooling/node-webgpu.d.ts +0 -17
- package/src/tooling/precision-replay-math.js +0 -174
- package/src/tooling/program-bundle-host.d.ts +0 -16
- package/src/tooling/program-bundle-host.js +0 -20
- package/src/tooling/runtime-input-composition.d.ts +0 -37
- package/src/tooling/source-package-profiles.d.ts +0 -50
- package/src/tooling/source-runtime-converter-config.d.ts +0 -19
- package/src/tooling/source-runtime-materializer.d.ts +0 -6
- package/src/tooling/source-runtime-materializer.js +0 -89
- package/src/tooling-exports.browser.d.ts +0 -7
- package/src/types/chrome.d.ts +0 -36
- package/src/types/chrome.js +0 -1
- package/src/types/index.d.ts +0 -1
- package/src/types/index.js +0 -1
- package/src/utils/array-buffer.d.ts +0 -4
- package/src/utils/canonical-hash.d.ts +0 -7
- package/src/utils/clone-json.d.ts +0 -1
- package/src/utils/hf-resolve-url.d.ts +0 -16
- package/src/utils/index.d.ts +0 -10
- package/src/utils/index.js +0 -10
- package/src/utils/load-json.d.ts +0 -5
- package/src/utils/plain-object.d.ts +0 -1
- package/src/utils/sha256.d.ts +0 -5
- package/src/utils/stable-sort-object.d.ts +0 -1
|
@@ -0,0 +1,208 @@
|
|
|
1
|
+
import { getRuntimeConfig } from '../../../config/runtime.js';
|
|
2
|
+
import { getDevice } from '../../../gpu/device.js';
|
|
3
|
+
import { createWeightBuffer, isWeightBuffer } from '../../../gpu/weight-buffer.js';
|
|
4
|
+
import { Q4K_BLOCK_BYTES, QK_K } from '../../../config/schema/index.js';
|
|
5
|
+
|
|
6
|
+
function shapeMatches(weight, rows, cols) {
|
|
7
|
+
return Array.isArray(weight?.shape)
|
|
8
|
+
&& weight.shape.length === 2
|
|
9
|
+
&& Number(weight.shape[0]) === rows
|
|
10
|
+
&& Number(weight.shape[1]) === cols;
|
|
11
|
+
}
|
|
12
|
+
|
|
13
|
+
function canUseLinearAttentionABProjectionFusion({
|
|
14
|
+
phase,
|
|
15
|
+
numTokens,
|
|
16
|
+
debugProbes,
|
|
17
|
+
operatorDiagnostics,
|
|
18
|
+
}) {
|
|
19
|
+
const session = getRuntimeConfig()?.inference?.session;
|
|
20
|
+
if (session?.useLinearAttentionABProjectionFusion !== true) {
|
|
21
|
+
return false;
|
|
22
|
+
}
|
|
23
|
+
if (phase !== 'decode' || numTokens !== 1) {
|
|
24
|
+
return false;
|
|
25
|
+
}
|
|
26
|
+
if (debugProbes?.length || operatorDiagnostics?.enabled === true) {
|
|
27
|
+
return false;
|
|
28
|
+
}
|
|
29
|
+
return true;
|
|
30
|
+
}
|
|
31
|
+
|
|
32
|
+
export function resolveLinearAttentionABProjection(layerWeights, options) {
|
|
33
|
+
const {
|
|
34
|
+
phase,
|
|
35
|
+
numTokens,
|
|
36
|
+
hiddenSize,
|
|
37
|
+
numVHeads,
|
|
38
|
+
layerIdx,
|
|
39
|
+
debugProbes,
|
|
40
|
+
operatorDiagnostics,
|
|
41
|
+
} = options;
|
|
42
|
+
if (!canUseLinearAttentionABProjectionFusion({
|
|
43
|
+
phase,
|
|
44
|
+
numTokens,
|
|
45
|
+
debugProbes,
|
|
46
|
+
operatorDiagnostics,
|
|
47
|
+
})) {
|
|
48
|
+
return null;
|
|
49
|
+
}
|
|
50
|
+
if (layerWeights.linearABProj && shapeMatches(layerWeights.linearABProj, numVHeads * 2, hiddenSize)) {
|
|
51
|
+
return {
|
|
52
|
+
weight: layerWeights.linearABProj,
|
|
53
|
+
outDim: numVHeads * 2,
|
|
54
|
+
bProjOffsetElements: numTokens * numVHeads,
|
|
55
|
+
};
|
|
56
|
+
}
|
|
57
|
+
|
|
58
|
+
const aWeight = layerWeights.linearInProjA;
|
|
59
|
+
const bWeight = layerWeights.linearInProjB;
|
|
60
|
+
if (
|
|
61
|
+
!isWeightBuffer(aWeight)
|
|
62
|
+
|| !isWeightBuffer(bWeight)
|
|
63
|
+
|| aWeight.dtype !== 'f16'
|
|
64
|
+
|| bWeight.dtype !== 'f16'
|
|
65
|
+
|| aWeight.layout !== 'row'
|
|
66
|
+
|| bWeight.layout !== 'row'
|
|
67
|
+
|| !shapeMatches(aWeight, numVHeads, hiddenSize)
|
|
68
|
+
|| !shapeMatches(bWeight, numVHeads, hiddenSize)
|
|
69
|
+
) {
|
|
70
|
+
return null;
|
|
71
|
+
}
|
|
72
|
+
|
|
73
|
+
const bytesPerWeight = numVHeads * hiddenSize * 2;
|
|
74
|
+
if (aWeight.buffer.size < bytesPerWeight || bWeight.buffer.size < bytesPerWeight) {
|
|
75
|
+
return null;
|
|
76
|
+
}
|
|
77
|
+
const device = getDevice();
|
|
78
|
+
if (!device) {
|
|
79
|
+
return null;
|
|
80
|
+
}
|
|
81
|
+
|
|
82
|
+
const buffer = device.createBuffer({
|
|
83
|
+
label: `L${layerIdx}.linear_ab_proj_weight`,
|
|
84
|
+
size: bytesPerWeight * 2,
|
|
85
|
+
usage: GPUBufferUsage.STORAGE | GPUBufferUsage.COPY_DST,
|
|
86
|
+
});
|
|
87
|
+
const encoder = device.createCommandEncoder({ label: `L${layerIdx}.linear_ab_proj_weight_pack` });
|
|
88
|
+
encoder.copyBufferToBuffer(aWeight.buffer, 0, buffer, 0, bytesPerWeight);
|
|
89
|
+
encoder.copyBufferToBuffer(bWeight.buffer, 0, buffer, bytesPerWeight, bytesPerWeight);
|
|
90
|
+
device.queue.submit([encoder.finish()]);
|
|
91
|
+
|
|
92
|
+
layerWeights.linearABProj = createWeightBuffer(
|
|
93
|
+
buffer,
|
|
94
|
+
'f16',
|
|
95
|
+
'row',
|
|
96
|
+
[numVHeads * 2, hiddenSize],
|
|
97
|
+
`L${layerIdx}.linear_ab_proj_weight`
|
|
98
|
+
);
|
|
99
|
+
return {
|
|
100
|
+
weight: layerWeights.linearABProj,
|
|
101
|
+
outDim: numVHeads * 2,
|
|
102
|
+
bProjOffsetElements: numTokens * numVHeads,
|
|
103
|
+
};
|
|
104
|
+
}
|
|
105
|
+
|
|
106
|
+
function canUseLinearAttentionQKVZProjectionFusion({
|
|
107
|
+
phase,
|
|
108
|
+
numTokens,
|
|
109
|
+
debugProbes,
|
|
110
|
+
operatorDiagnostics,
|
|
111
|
+
}) {
|
|
112
|
+
const session = getRuntimeConfig()?.inference?.session;
|
|
113
|
+
if (session?.useLinearAttentionQKVZProjectionFusion !== true) {
|
|
114
|
+
return false;
|
|
115
|
+
}
|
|
116
|
+
if (phase !== 'decode' || numTokens !== 1) {
|
|
117
|
+
return false;
|
|
118
|
+
}
|
|
119
|
+
if (debugProbes?.length || operatorDiagnostics?.enabled === true) {
|
|
120
|
+
return false;
|
|
121
|
+
}
|
|
122
|
+
return true;
|
|
123
|
+
}
|
|
124
|
+
|
|
125
|
+
function q4kRowBytes(hiddenSize) {
|
|
126
|
+
if (!Number.isInteger(hiddenSize) || hiddenSize <= 0) {
|
|
127
|
+
return null;
|
|
128
|
+
}
|
|
129
|
+
return Math.ceil(hiddenSize / QK_K) * Q4K_BLOCK_BYTES;
|
|
130
|
+
}
|
|
131
|
+
|
|
132
|
+
export function resolveLinearAttentionQKVZProjection(layerWeights, options) {
|
|
133
|
+
const {
|
|
134
|
+
phase,
|
|
135
|
+
numTokens,
|
|
136
|
+
hiddenSize,
|
|
137
|
+
convDim,
|
|
138
|
+
valueDim,
|
|
139
|
+
layerIdx,
|
|
140
|
+
debugProbes,
|
|
141
|
+
operatorDiagnostics,
|
|
142
|
+
} = options;
|
|
143
|
+
if (!canUseLinearAttentionQKVZProjectionFusion({
|
|
144
|
+
phase,
|
|
145
|
+
numTokens,
|
|
146
|
+
debugProbes,
|
|
147
|
+
operatorDiagnostics,
|
|
148
|
+
})) {
|
|
149
|
+
return null;
|
|
150
|
+
}
|
|
151
|
+
if (layerWeights.linearQKVZProj && shapeMatches(layerWeights.linearQKVZProj, convDim + valueDim, hiddenSize)) {
|
|
152
|
+
return {
|
|
153
|
+
weight: layerWeights.linearQKVZProj,
|
|
154
|
+
outDim: convDim + valueDim,
|
|
155
|
+
};
|
|
156
|
+
}
|
|
157
|
+
|
|
158
|
+
const qkvWeight = layerWeights.qkvProj;
|
|
159
|
+
const zWeight = layerWeights.linearInProjZ;
|
|
160
|
+
if (
|
|
161
|
+
!isWeightBuffer(qkvWeight)
|
|
162
|
+
|| !isWeightBuffer(zWeight)
|
|
163
|
+
|| qkvWeight.dtype !== 'q4k'
|
|
164
|
+
|| zWeight.dtype !== 'q4k'
|
|
165
|
+
|| qkvWeight.layout !== 'row'
|
|
166
|
+
|| zWeight.layout !== 'row'
|
|
167
|
+
|| !shapeMatches(qkvWeight, convDim, hiddenSize)
|
|
168
|
+
|| !shapeMatches(zWeight, valueDim, hiddenSize)
|
|
169
|
+
) {
|
|
170
|
+
return null;
|
|
171
|
+
}
|
|
172
|
+
|
|
173
|
+
const rowBytes = q4kRowBytes(hiddenSize);
|
|
174
|
+
if (rowBytes == null) {
|
|
175
|
+
return null;
|
|
176
|
+
}
|
|
177
|
+
const qkvBytes = convDim * rowBytes;
|
|
178
|
+
const zBytes = valueDim * rowBytes;
|
|
179
|
+
if (qkvWeight.buffer.size < qkvBytes || zWeight.buffer.size < zBytes) {
|
|
180
|
+
return null;
|
|
181
|
+
}
|
|
182
|
+
const device = getDevice();
|
|
183
|
+
if (!device) {
|
|
184
|
+
return null;
|
|
185
|
+
}
|
|
186
|
+
|
|
187
|
+
const buffer = device.createBuffer({
|
|
188
|
+
label: `L${layerIdx}.linear_qkvz_proj_weight`,
|
|
189
|
+
size: qkvBytes + zBytes,
|
|
190
|
+
usage: GPUBufferUsage.STORAGE | GPUBufferUsage.COPY_DST,
|
|
191
|
+
});
|
|
192
|
+
const encoder = device.createCommandEncoder({ label: `L${layerIdx}.linear_qkvz_proj_weight_pack` });
|
|
193
|
+
encoder.copyBufferToBuffer(qkvWeight.buffer, 0, buffer, 0, qkvBytes);
|
|
194
|
+
encoder.copyBufferToBuffer(zWeight.buffer, 0, buffer, qkvBytes, zBytes);
|
|
195
|
+
device.queue.submit([encoder.finish()]);
|
|
196
|
+
|
|
197
|
+
layerWeights.linearQKVZProj = createWeightBuffer(
|
|
198
|
+
buffer,
|
|
199
|
+
'q4k',
|
|
200
|
+
'row',
|
|
201
|
+
[convDim + valueDim, hiddenSize],
|
|
202
|
+
`L${layerIdx}.linear_qkvz_proj_weight`
|
|
203
|
+
);
|
|
204
|
+
return {
|
|
205
|
+
weight: layerWeights.linearQKVZProj,
|
|
206
|
+
outDim: convDim + valueDim,
|
|
207
|
+
};
|
|
208
|
+
}
|
|
@@ -73,6 +73,7 @@ export interface RunLinearAttentionLayerOptions {
|
|
|
73
73
|
weight: GPUBuffer | Float32Array | ArrayBuffer,
|
|
74
74
|
label: string
|
|
75
75
|
) => GPUBuffer;
|
|
76
|
+
precomputedInputNorm?: Tensor | null;
|
|
76
77
|
debugProbes?: ProbeConfigSchema[] | null;
|
|
77
78
|
recorder?: CommandRecorder | null;
|
|
78
79
|
}
|
|
@@ -1,9 +1,22 @@
|
|
|
1
1
|
import { getBufferDtype, isGpuBufferInstance, isWeightBuffer } from '../../../gpu/weight-buffer.js';
|
|
2
|
-
import {
|
|
2
|
+
import {
|
|
3
|
+
recordMatmul,
|
|
4
|
+
recordRMSNorm,
|
|
5
|
+
runMatmul,
|
|
6
|
+
runRMSNorm,
|
|
7
|
+
castF32ToF16,
|
|
8
|
+
castF16ToF32,
|
|
9
|
+
recordCastF32ToF16,
|
|
10
|
+
recordCastF16ToF32,
|
|
11
|
+
} from '../../../gpu/kernel-selector.js';
|
|
3
12
|
import { readBuffer, releaseBuffer, uploadData, acquireBuffer } from '../../../memory/buffer-pool.js';
|
|
4
13
|
import { log } from '../../../debug/index.js';
|
|
5
14
|
import { decodeReadback, f16ToF32 } from './debug-utils/index.js';
|
|
6
15
|
import { runLinearAttentionCoreGPU } from '../../../gpu/kernels/linear-attention-core.js';
|
|
16
|
+
import {
|
|
17
|
+
resolveLinearAttentionABProjection,
|
|
18
|
+
resolveLinearAttentionQKVZProjection,
|
|
19
|
+
} from './linear-attention-ab-fusion.js';
|
|
7
20
|
import { runProbes } from './probes.js';
|
|
8
21
|
import { QK_K, Q4K_BLOCK_BYTES } from '../../../config/schema/index.js';
|
|
9
22
|
import { dequantizeQ4KM } from '../../../converter/quantizer.js';
|
|
@@ -656,9 +669,32 @@ async function projectLinearTensor({
|
|
|
656
669
|
? await recordCastF32ToF16(recorder, inputTensor)
|
|
657
670
|
: await castF32ToF16(inputTensor);
|
|
658
671
|
}
|
|
672
|
+
const settleProjectionOutputDtype = async (result) => {
|
|
673
|
+
if (result.dtype === resolvedOutputDtype) {
|
|
674
|
+
return result;
|
|
675
|
+
}
|
|
676
|
+
assertImplicitDtypeTransitionAllowed({
|
|
677
|
+
executionPolicies,
|
|
678
|
+
fromDtype: result.dtype,
|
|
679
|
+
toDtype: resolvedOutputDtype,
|
|
680
|
+
op: role,
|
|
681
|
+
detail: 'Linear attention projection returned a kernel-selected dtype different from the declared projection output dtype.',
|
|
682
|
+
});
|
|
683
|
+
if (resolvedOutputDtype === 'f16') {
|
|
684
|
+
const casted = recorder ? await recordCastF32ToF16(recorder, result) : await castF32ToF16(result);
|
|
685
|
+
releaseOrTrackBuffer(recorder, result.buffer);
|
|
686
|
+
return casted;
|
|
687
|
+
}
|
|
688
|
+
if (resolvedOutputDtype === 'f32') {
|
|
689
|
+
const casted = recorder ? await recordCastF16ToF32(recorder, result) : await castF16ToF32(result);
|
|
690
|
+
releaseOrTrackBuffer(recorder, result.buffer);
|
|
691
|
+
return casted;
|
|
692
|
+
}
|
|
693
|
+
throw new Error(`Unsupported linear_attention projection output dtype "${String(resolvedOutputDtype)}".`);
|
|
694
|
+
};
|
|
659
695
|
try {
|
|
660
696
|
if (recorder) {
|
|
661
|
-
|
|
697
|
+
const result = await recordMatmul(recorder, matmulInput, resolvedWeight, numTokens, outDim, hiddenSize, {
|
|
662
698
|
transposeB: 'auto',
|
|
663
699
|
role,
|
|
664
700
|
layerIdx,
|
|
@@ -666,8 +702,9 @@ async function projectLinearTensor({
|
|
|
666
702
|
outputDtype: resolvedOutputDtype,
|
|
667
703
|
executionPolicies,
|
|
668
704
|
});
|
|
705
|
+
return await settleProjectionOutputDtype(result);
|
|
669
706
|
}
|
|
670
|
-
|
|
707
|
+
const result = await runMatmul(matmulInput, resolvedWeight, numTokens, outDim, hiddenSize, {
|
|
671
708
|
transposeB: 'auto',
|
|
672
709
|
role,
|
|
673
710
|
layerIdx,
|
|
@@ -675,6 +712,7 @@ async function projectLinearTensor({
|
|
|
675
712
|
outputDtype: resolvedOutputDtype,
|
|
676
713
|
executionPolicies,
|
|
677
714
|
});
|
|
715
|
+
return await settleProjectionOutputDtype(result);
|
|
678
716
|
} finally {
|
|
679
717
|
if (matmulInput !== inputTensor) {
|
|
680
718
|
releaseOrTrackBuffer(recorder, matmulInput.buffer);
|
|
@@ -683,6 +721,31 @@ async function projectLinearTensor({
|
|
|
683
721
|
}
|
|
684
722
|
}
|
|
685
723
|
|
|
724
|
+
async function settleLinearAttentionCoreInputDtype(tensor, targetDtype, options) {
|
|
725
|
+
if (tensor.dtype === targetDtype) {
|
|
726
|
+
return tensor;
|
|
727
|
+
}
|
|
728
|
+
const {
|
|
729
|
+
recorder,
|
|
730
|
+
executionPolicies,
|
|
731
|
+
role,
|
|
732
|
+
} = options;
|
|
733
|
+
assertImplicitDtypeTransitionAllowed({
|
|
734
|
+
executionPolicies,
|
|
735
|
+
fromDtype: tensor.dtype,
|
|
736
|
+
toDtype: targetDtype,
|
|
737
|
+
op: role,
|
|
738
|
+
detail: 'Linear attention core requires qkv/z/a/b tensors to share one input dtype.',
|
|
739
|
+
});
|
|
740
|
+
if (targetDtype === 'f16') {
|
|
741
|
+
return recorder ? await recordCastF32ToF16(recorder, tensor) : await castF32ToF16(tensor);
|
|
742
|
+
}
|
|
743
|
+
if (targetDtype === 'f32') {
|
|
744
|
+
return recorder ? await recordCastF16ToF32(recorder, tensor) : await castF16ToF32(tensor);
|
|
745
|
+
}
|
|
746
|
+
throw new Error(`Unsupported linear_attention core input dtype "${String(targetDtype)}".`);
|
|
747
|
+
}
|
|
748
|
+
|
|
686
749
|
export function hasLinearAttentionLayers(layerTypes) {
|
|
687
750
|
if (!Array.isArray(layerTypes)) return false;
|
|
688
751
|
for (let i = 0; i < layerTypes.length; i++) {
|
|
@@ -771,6 +834,7 @@ export async function runLinearAttentionLayer(inputTensor, layerWeights, options
|
|
|
771
834
|
getNormWeightBuffer,
|
|
772
835
|
recorder,
|
|
773
836
|
executionPolicies = null,
|
|
837
|
+
precomputedInputNorm = null,
|
|
774
838
|
} = options;
|
|
775
839
|
|
|
776
840
|
if (!layerWeights) {
|
|
@@ -808,7 +872,20 @@ export async function runLinearAttentionLayer(inputTensor, layerWeights, options
|
|
|
808
872
|
let normedTensor = inputTensor;
|
|
809
873
|
let normedCreated = false;
|
|
810
874
|
|
|
811
|
-
if (
|
|
875
|
+
if (precomputedInputNorm) {
|
|
876
|
+
if (
|
|
877
|
+
phase !== 'decode'
|
|
878
|
+
|| numTokens !== 1
|
|
879
|
+
|| !layerWeights.inputNorm
|
|
880
|
+
|| precomputedInputNorm.dtype !== inputTensor.dtype
|
|
881
|
+
|| precomputedInputNorm.dtype !== 'f32'
|
|
882
|
+
) {
|
|
883
|
+
releaseOrTrackBuffer(recorder, precomputedInputNorm.buffer);
|
|
884
|
+
throw new Error(`linear_attention layer ${layerIdx} received an incompatible precomputed input norm tensor.`);
|
|
885
|
+
}
|
|
886
|
+
normedTensor = precomputedInputNorm;
|
|
887
|
+
normedCreated = true;
|
|
888
|
+
} else if (layerWeights.inputNorm) {
|
|
812
889
|
const normWeightBuffer = getNormWeightBuffer(layerWeights.inputNorm, `L${layerIdx}.linear_input_norm`);
|
|
813
890
|
try {
|
|
814
891
|
if (recorder) {
|
|
@@ -832,7 +909,35 @@ export async function runLinearAttentionLayer(inputTensor, layerWeights, options
|
|
|
832
909
|
}
|
|
833
910
|
}
|
|
834
911
|
|
|
835
|
-
const
|
|
912
|
+
const qkvzProjection = resolveLinearAttentionQKVZProjection(layerWeights, {
|
|
913
|
+
phase,
|
|
914
|
+
numTokens,
|
|
915
|
+
hiddenSize,
|
|
916
|
+
numVHeads: projectionLayout.numVHeads,
|
|
917
|
+
convDim: projectionLayout.convDim,
|
|
918
|
+
valueDim: projectionLayout.valueDim,
|
|
919
|
+
layerIdx,
|
|
920
|
+
debugProbes: options.debugProbes,
|
|
921
|
+
operatorDiagnostics: options.operatorDiagnostics,
|
|
922
|
+
});
|
|
923
|
+
const qkvzTensor = qkvzProjection
|
|
924
|
+
? await projectLinearTensor({
|
|
925
|
+
inputTensor: normedTensor,
|
|
926
|
+
sourceWeight: qkvzProjection.weight,
|
|
927
|
+
role: 'linear_qkvz_proj',
|
|
928
|
+
phase,
|
|
929
|
+
outDim: qkvzProjection.outDim,
|
|
930
|
+
numTokens,
|
|
931
|
+
hiddenSize,
|
|
932
|
+
layerIdx,
|
|
933
|
+
kernelPath,
|
|
934
|
+
outputDtype: projectionDtype,
|
|
935
|
+
getWeightBuffer,
|
|
936
|
+
recorder,
|
|
937
|
+
executionPolicies,
|
|
938
|
+
})
|
|
939
|
+
: null;
|
|
940
|
+
const qkvTensor = qkvzTensor ?? await projectLinearTensor({
|
|
836
941
|
inputTensor: normedTensor,
|
|
837
942
|
sourceWeight: layerWeights.qkvProj,
|
|
838
943
|
role: 'linear_qkv_proj',
|
|
@@ -847,7 +952,7 @@ export async function runLinearAttentionLayer(inputTensor, layerWeights, options
|
|
|
847
952
|
recorder,
|
|
848
953
|
executionPolicies,
|
|
849
954
|
});
|
|
850
|
-
const zTensor = await projectLinearTensor({
|
|
955
|
+
const zTensor = qkvzTensor ?? await projectLinearTensor({
|
|
851
956
|
inputTensor: normedTensor,
|
|
852
957
|
sourceWeight: layerWeights.linearInProjZ,
|
|
853
958
|
role: 'linear_z_proj',
|
|
@@ -862,7 +967,33 @@ export async function runLinearAttentionLayer(inputTensor, layerWeights, options
|
|
|
862
967
|
recorder,
|
|
863
968
|
executionPolicies,
|
|
864
969
|
});
|
|
865
|
-
const
|
|
970
|
+
const abProjection = resolveLinearAttentionABProjection(layerWeights, {
|
|
971
|
+
phase,
|
|
972
|
+
numTokens,
|
|
973
|
+
hiddenSize,
|
|
974
|
+
numVHeads: projectionLayout.numVHeads,
|
|
975
|
+
layerIdx,
|
|
976
|
+
debugProbes: options.debugProbes,
|
|
977
|
+
operatorDiagnostics: options.operatorDiagnostics,
|
|
978
|
+
});
|
|
979
|
+
const abTensor = abProjection
|
|
980
|
+
? await projectLinearTensor({
|
|
981
|
+
inputTensor: normedTensor,
|
|
982
|
+
sourceWeight: abProjection.weight,
|
|
983
|
+
role: 'linear_ab_proj',
|
|
984
|
+
phase,
|
|
985
|
+
outDim: abProjection.outDim,
|
|
986
|
+
numTokens,
|
|
987
|
+
hiddenSize,
|
|
988
|
+
layerIdx,
|
|
989
|
+
kernelPath,
|
|
990
|
+
outputDtype: projectionDtype,
|
|
991
|
+
getWeightBuffer,
|
|
992
|
+
recorder,
|
|
993
|
+
executionPolicies,
|
|
994
|
+
})
|
|
995
|
+
: null;
|
|
996
|
+
const aTensor = abTensor ?? await projectLinearTensor({
|
|
866
997
|
inputTensor: normedTensor,
|
|
867
998
|
sourceWeight: layerWeights.linearInProjA,
|
|
868
999
|
role: 'linear_a_proj',
|
|
@@ -877,7 +1008,7 @@ export async function runLinearAttentionLayer(inputTensor, layerWeights, options
|
|
|
877
1008
|
recorder,
|
|
878
1009
|
executionPolicies,
|
|
879
1010
|
});
|
|
880
|
-
const bTensor = await projectLinearTensor({
|
|
1011
|
+
const bTensor = abTensor ?? await projectLinearTensor({
|
|
881
1012
|
inputTensor: normedTensor,
|
|
882
1013
|
sourceWeight: layerWeights.linearInProjB,
|
|
883
1014
|
role: 'linear_b_proj',
|
|
@@ -909,6 +1040,10 @@ export async function runLinearAttentionLayer(inputTensor, layerWeights, options
|
|
|
909
1040
|
layerOutputDtype,
|
|
910
1041
|
'outputDtype'
|
|
911
1042
|
);
|
|
1043
|
+
let coreQkvTensor = qkvTensor;
|
|
1044
|
+
let coreZTensor = zTensor;
|
|
1045
|
+
let coreATensor = aTensor;
|
|
1046
|
+
let coreBTensor = bTensor;
|
|
912
1047
|
|
|
913
1048
|
try {
|
|
914
1049
|
await runProbes('linear_qkv_proj', qkvTensor.buffer, {
|
|
@@ -947,11 +1082,27 @@ export async function runLinearAttentionLayer(inputTensor, layerWeights, options
|
|
|
947
1082
|
operatorDiagnostics: options.operatorDiagnostics,
|
|
948
1083
|
dtype: bTensor.dtype,
|
|
949
1084
|
});
|
|
1085
|
+
const coreInputDtype = qkvTensor.dtype;
|
|
1086
|
+
coreZTensor = await settleLinearAttentionCoreInputDtype(zTensor, coreInputDtype, {
|
|
1087
|
+
recorder,
|
|
1088
|
+
executionPolicies,
|
|
1089
|
+
role: 'linear_z_proj',
|
|
1090
|
+
});
|
|
1091
|
+
coreATensor = await settleLinearAttentionCoreInputDtype(aTensor, coreInputDtype, {
|
|
1092
|
+
recorder,
|
|
1093
|
+
executionPolicies,
|
|
1094
|
+
role: 'linear_a_proj',
|
|
1095
|
+
});
|
|
1096
|
+
coreBTensor = await settleLinearAttentionCoreInputDtype(bTensor, coreInputDtype, {
|
|
1097
|
+
recorder,
|
|
1098
|
+
executionPolicies,
|
|
1099
|
+
role: 'linear_b_proj',
|
|
1100
|
+
});
|
|
950
1101
|
const coreTensor = await runLinearAttentionCoreGPU(
|
|
951
|
-
|
|
952
|
-
|
|
953
|
-
|
|
954
|
-
|
|
1102
|
+
coreQkvTensor,
|
|
1103
|
+
coreZTensor,
|
|
1104
|
+
coreATensor,
|
|
1105
|
+
coreBTensor,
|
|
955
1106
|
layerState,
|
|
956
1107
|
{
|
|
957
1108
|
numTokens,
|
|
@@ -960,6 +1111,9 @@ export async function runLinearAttentionLayer(inputTensor, layerWeights, options
|
|
|
960
1111
|
qkL2NormEps: QK_L2NORM_EPS,
|
|
961
1112
|
recorder,
|
|
962
1113
|
executionPolicies,
|
|
1114
|
+
abPacked: abProjection != null,
|
|
1115
|
+
qkvzPacked: qkvzProjection != null,
|
|
1116
|
+
bProjOffsetElements: abProjection?.bProjOffsetElements ?? 0,
|
|
963
1117
|
}
|
|
964
1118
|
);
|
|
965
1119
|
await runProbes('linear_core_out', coreTensor.buffer, {
|
|
@@ -1017,9 +1171,12 @@ export async function runLinearAttentionLayer(inputTensor, layerWeights, options
|
|
|
1017
1171
|
if (normedCreated) {
|
|
1018
1172
|
releaseOrTrackBuffer(recorder, normedTensor.buffer);
|
|
1019
1173
|
}
|
|
1020
|
-
|
|
1021
|
-
|
|
1022
|
-
|
|
1023
|
-
|
|
1174
|
+
const released = new Set();
|
|
1175
|
+
for (const tensor of [qkvTensor, zTensor, aTensor, bTensor, coreQkvTensor, coreZTensor, coreATensor, coreBTensor]) {
|
|
1176
|
+
const buffer = tensor?.buffer ?? null;
|
|
1177
|
+
if (!buffer || released.has(buffer)) continue;
|
|
1178
|
+
released.add(buffer);
|
|
1179
|
+
releaseOrTrackBuffer(recorder, buffer);
|
|
1180
|
+
}
|
|
1024
1181
|
}
|
|
1025
1182
|
}
|
|
@@ -5,7 +5,7 @@ import { acquireBuffer, releaseBuffer } from '../../../../memory/buffer-pool.js'
|
|
|
5
5
|
import { runMatmul, runRMSNorm, runScale, recordScale } from '../../../../gpu/kernel-selector.js';
|
|
6
6
|
import { recordMatmul } from '../../../../gpu/kernels/matmul.js';
|
|
7
7
|
import { recordRMSNorm } from '../../../../gpu/kernels/rmsnorm.js';
|
|
8
|
-
import {
|
|
8
|
+
import { recordLmHeadArgmax } from '../../../../gpu/kernels/lm-head-argmax.js';
|
|
9
9
|
import { createTensor } from '../../../../gpu/tensor.js';
|
|
10
10
|
import {
|
|
11
11
|
castF16ToF32,
|
|
@@ -1176,7 +1176,7 @@ export async function recordGreedyLmHeadArgmaxGPU(
|
|
|
1176
1176
|
config,
|
|
1177
1177
|
operatorDiagnostics
|
|
1178
1178
|
);
|
|
1179
|
-
const outputBuffer = await
|
|
1179
|
+
const outputBuffer = await recordLmHeadArgmax(recorder, tail.lmHeadInputTensor, tail.lmHeadBuffer, {
|
|
1180
1180
|
vocabSize: tail.matmulVocabSize,
|
|
1181
1181
|
hiddenSize: tail.hiddenSize,
|
|
1182
1182
|
padTokenId: options.padTokenId,
|
|
@@ -12,7 +12,7 @@ export { extractLastPositionLogits, finalizeLogits, readBufferWithCleanup } from
|
|
|
12
12
|
// Imports for computeLogits orchestrator
|
|
13
13
|
import { getDevice } from '../../../../gpu/device.js';
|
|
14
14
|
import { acquireBuffer, releaseBuffer, readBuffer } from '../../../../memory/buffer-pool.js';
|
|
15
|
-
import { runMatmul, runRMSNorm, runScale, castF16ToF32, castF32ToF16 } from '../../../../gpu/kernel-selector.js';
|
|
15
|
+
import { runMatmul, runRMSNorm, runScale, castF16ToF32, castF32ToF16, runLmHeadSelectLogitsF16 } from '../../../../gpu/kernel-selector.js';
|
|
16
16
|
import { createTensor } from '../../../../gpu/tensor.js';
|
|
17
17
|
import { isWeightBuffer, isCpuWeightBuffer, isGpuBufferInstance, isSplitWeightBuffer, getWeightDtype } from '../../../../gpu/weight-buffer.js';
|
|
18
18
|
import { kernelTrace, traceStep } from '../kernel-trace.js';
|
|
@@ -50,6 +50,28 @@ function resolveLmHeadMatmulRole(phase) {
|
|
|
50
50
|
return phase === 'prefill' ? 'lm_head_prefill' : 'lm_head';
|
|
51
51
|
}
|
|
52
52
|
|
|
53
|
+
function normalizeSelectedLogitTokenIds(value, vocabSize) {
|
|
54
|
+
if (value == null) {
|
|
55
|
+
return null;
|
|
56
|
+
}
|
|
57
|
+
if (!Array.isArray(value) && !ArrayBuffer.isView(value)) {
|
|
58
|
+
throw new Error('[Logits] selectedTokenIds must be an array or typed array.');
|
|
59
|
+
}
|
|
60
|
+
const tokenIds = Array.from(value, (entry, index) => {
|
|
61
|
+
const tokenId = Number(entry);
|
|
62
|
+
if (!Number.isInteger(tokenId) || tokenId < 0 || tokenId >= vocabSize) {
|
|
63
|
+
throw new Error(
|
|
64
|
+
`[Logits] selectedTokenIds[${index}] must be an integer in [0, ${vocabSize}), got "${String(entry)}".`
|
|
65
|
+
);
|
|
66
|
+
}
|
|
67
|
+
return tokenId;
|
|
68
|
+
});
|
|
69
|
+
if (tokenIds.length === 0) {
|
|
70
|
+
throw new Error('[Logits] selectedTokenIds must not be empty.');
|
|
71
|
+
}
|
|
72
|
+
return tokenIds;
|
|
73
|
+
}
|
|
74
|
+
|
|
53
75
|
async function coerceTensorDtype(tensor, targetDtype, options = {}) {
|
|
54
76
|
if (!targetDtype || tensor.dtype === targetDtype) {
|
|
55
77
|
return tensor;
|
|
@@ -196,6 +218,7 @@ export async function computeLogits(
|
|
|
196
218
|
matmulVocabSize = dims.vocabSize;
|
|
197
219
|
}
|
|
198
220
|
}
|
|
221
|
+
const selectedTokenIds = normalizeSelectedLogitTokenIds(options?.selectedTokenIds, matmulVocabSize);
|
|
199
222
|
|
|
200
223
|
// Check if input is GPU buffer
|
|
201
224
|
const inputIsGPU = isGpuBufferInstance(hiddenStates);
|
|
@@ -204,6 +227,12 @@ export async function computeLogits(
|
|
|
204
227
|
if (isTraceEnabled('logits')) {
|
|
205
228
|
trace.logits(`LOGITS_PATH: device=${!!device}, useGPU=${useGPU}, taking ${(!device || !useGPU) ? 'CPU' : 'GPU'} path`);
|
|
206
229
|
}
|
|
230
|
+
if (selectedTokenIds && (!device || !useGPU)) {
|
|
231
|
+
throw new Error('[Logits] selectedTokenIds requires the GPU logits path.');
|
|
232
|
+
}
|
|
233
|
+
if (selectedTokenIds && (isCpuWeightBuffer(lmHead) || isSplitWeightBuffer(lmHead))) {
|
|
234
|
+
throw new Error('[Logits] selectedTokenIds requires a single GPU-resident LM head weight buffer.');
|
|
235
|
+
}
|
|
207
236
|
if (!device || !useGPU) {
|
|
208
237
|
|
|
209
238
|
let cpuHiddenStates;
|
|
@@ -481,6 +510,38 @@ export async function computeLogits(
|
|
|
481
510
|
matmulInputOwned = true;
|
|
482
511
|
}
|
|
483
512
|
|
|
513
|
+
if (selectedTokenIds) {
|
|
514
|
+
if (lastPositionOnly !== true) {
|
|
515
|
+
throw new Error('[Logits] selectedTokenIds requires lastPositionOnly=true.');
|
|
516
|
+
}
|
|
517
|
+
const selected = await runLmHeadSelectLogitsF16(matmulInputTensor, lmHeadBuffer, {
|
|
518
|
+
device,
|
|
519
|
+
hiddenSize,
|
|
520
|
+
vocabSize: matmulVocabSize,
|
|
521
|
+
tokenIds: selectedTokenIds,
|
|
522
|
+
hiddenOffset: 0,
|
|
523
|
+
logitSoftcap: config.finalLogitSoftcapping == null ? 0 : Number(config.finalLogitSoftcapping),
|
|
524
|
+
});
|
|
525
|
+
const selectedBytes = selectedTokenIds.length * Float32Array.BYTES_PER_ELEMENT;
|
|
526
|
+
const selectedData = await readBufferWithCleanup(selected.outputBuffer, selectedBytes, () => {
|
|
527
|
+
releaseBuffer(selected.outputBuffer);
|
|
528
|
+
releaseBuffer(selected.tokenIdBuffer);
|
|
529
|
+
if (inputBufferOwned) releaseBuffer(inputBuffer);
|
|
530
|
+
releaseBuffer(finalNormTensor.buffer);
|
|
531
|
+
if (matmulInputOwned) releaseBuffer(matmulInputTensor.buffer);
|
|
532
|
+
if (normWeightBufferOwned) releaseBuffer(normWeightBuffer);
|
|
533
|
+
if (lmHeadBufferOwned) releaseBuffer(lmHeadGPU);
|
|
534
|
+
});
|
|
535
|
+
const selectedLogits = new Float32Array(selectedData);
|
|
536
|
+
await runProbes('logits_final', selectedLogits, {
|
|
537
|
+
numTokens: 1,
|
|
538
|
+
hiddenSize: selectedTokenIds.length,
|
|
539
|
+
probes: debugProbes,
|
|
540
|
+
operatorDiagnostics,
|
|
541
|
+
});
|
|
542
|
+
return selectedLogits;
|
|
543
|
+
}
|
|
544
|
+
|
|
484
545
|
// HuggingFace models store lm_head as [vocabSize, hiddenSize], so transposeB=true
|
|
485
546
|
const logitsTensor = await runMatmul(matmulInputTensor, lmHeadBuffer, matmulRows, matmulVocabSize, hiddenSize, {
|
|
486
547
|
transposeB: 'auto',
|
|
@@ -155,6 +155,11 @@ function buildResolvedDecodeLoopRuntimePatch(
|
|
|
155
155
|
);
|
|
156
156
|
}
|
|
157
157
|
const submitLatencyThresholdMs = decodeLoop.submitLatencyThresholdMs ?? null;
|
|
158
|
+
const maxBatchDecodeTokens = parseSessionDecodeLoopOptionalPositiveInt(
|
|
159
|
+
decodeLoop.maxBatchDecodeTokens,
|
|
160
|
+
'maxBatchDecodeTokens',
|
|
161
|
+
modelId
|
|
162
|
+
);
|
|
158
163
|
|
|
159
164
|
const batchingPatch = {
|
|
160
165
|
batchSize,
|
|
@@ -196,6 +201,7 @@ function buildResolvedDecodeLoopRuntimePatch(
|
|
|
196
201
|
readbackInterval,
|
|
197
202
|
readbackMode,
|
|
198
203
|
submitLatencyThresholdMs,
|
|
204
|
+
...(maxBatchDecodeTokens !== undefined ? { maxBatchDecodeTokens } : {}),
|
|
199
205
|
...(ringTokens !== undefined ? { ringTokens } : {}),
|
|
200
206
|
...(ringStop !== undefined ? { ringStop } : {}),
|
|
201
207
|
...(ringStaging !== undefined ? { ringStaging } : {}),
|
|
@@ -5,7 +5,9 @@ import {
|
|
|
5
5
|
recordRMSNorm, recordResidualAdd, recordMatmul, recordSiLU, recordGeLU,
|
|
6
6
|
runSiLURowSplit, recordSiLURowSplit,
|
|
7
7
|
runMatmulRMSNormFused, recordMatmulRMSNormFused,
|
|
8
|
+
runRMSNormStats, recordRMSNormStats,
|
|
8
9
|
runSandwichRMSNormPair, recordSandwichRMSNormPair,
|
|
10
|
+
runResidualNextRMSNormPair, recordResidualNextRMSNormPair,
|
|
9
11
|
runConv2D, recordConv2D,
|
|
10
12
|
} from '../../../gpu/kernel-selector.js';
|
|
11
13
|
import {
|
|
@@ -58,6 +60,12 @@ export async function doRMSNorm(input, weight, eps, options, recorder) {
|
|
|
58
60
|
return result;
|
|
59
61
|
}
|
|
60
62
|
|
|
63
|
+
export async function doRMSNormStats(input, residual, eps, options, recorder) {
|
|
64
|
+
return recorder
|
|
65
|
+
? await recordRMSNormStats(recorder, input, residual, eps, options)
|
|
66
|
+
: await runRMSNormStats(input, residual, eps, options);
|
|
67
|
+
}
|
|
68
|
+
|
|
61
69
|
export async function doSandwichRMSNormPair(input, residual, postWeight, preWeight, eps, options, recorder) {
|
|
62
70
|
const result = recorder
|
|
63
71
|
? await recordSandwichRMSNormPair(recorder, input, residual, postWeight, preWeight, eps, options)
|
|
@@ -73,6 +81,20 @@ export async function doSandwichRMSNormPair(input, residual, postWeight, preWeig
|
|
|
73
81
|
return result;
|
|
74
82
|
}
|
|
75
83
|
|
|
84
|
+
export async function doResidualNextRMSNormPair(input, residual, normWeight, eps, options = {}, recorder) {
|
|
85
|
+
const result = recorder
|
|
86
|
+
? await recordResidualNextRMSNormPair(recorder, input, residual, normWeight, eps, options)
|
|
87
|
+
: await runResidualNextRMSNormPair(input, residual, normWeight, eps, options);
|
|
88
|
+
|
|
89
|
+
if (kernelTrace.enabled && !recorder) {
|
|
90
|
+
const layer = options.layerIdx ?? -1;
|
|
91
|
+
const label = options.label ?? 'residual_rmsnorm_pair';
|
|
92
|
+
await traceStep('residual_rmsnorm_pair.residual', label, layer, result.residual.buffer, [options.batchSize, options.hiddenSize]);
|
|
93
|
+
await traceStep('residual_rmsnorm_pair.next_input', label, layer, result.nextInput.buffer, [options.batchSize, options.hiddenSize]);
|
|
94
|
+
}
|
|
95
|
+
|
|
96
|
+
return result;
|
|
97
|
+
}
|
|
76
98
|
|
|
77
99
|
export async function doResidualAdd(a, b, size, recorder, traceOptions) {
|
|
78
100
|
const options = {};
|