doppler-gpu 0.4.6 → 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 +98 -137
- package/assets/doppler.svg +1 -3
- package/models/catalog.json +1575 -0
- package/package.json +59 -26
- package/src/cli/doppler-cli.js +28 -7
- package/src/cli/doppler-quickstart.js +1 -1
- package/src/cli/doppler-serve.js +2 -2
- package/src/client/doppler-registry.js +12 -0
- package/src/client/doppler-registry.json +52 -3
- 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/diffusiongemma/diffusiongemma-26b-a4b-it-q4k-ehf16-af16.json +28 -4
- package/src/config/conversion/embeddinggemma/google-embeddinggemma-300m-q4k-ehf16-af32.json +29 -4
- package/src/config/conversion/gemma3/gemma-3-1b-it-f16-af32.json +29 -4
- package/src/config/conversion/gemma3/gemma-3-1b-it-q4k-ehf16-af32.json +29 -4
- package/src/config/conversion/gemma3/gemma-3-270m-it-f16-af32.json +29 -4
- package/src/config/conversion/gemma3/gemma-3-270m-it-q4k-ehf16-af16.json +29 -4
- package/src/config/conversion/gemma3/gemma-3-270m-it-q4k-ehf16-af32.json +29 -4
- package/src/config/conversion/gemma3/translategemma-4b-1b-enes-q4k-ehf16-af32.json +50 -9
- package/src/config/conversion/gemma3/translategemma-4b-it-q4k-ehf16-af32.json +35 -5
- package/src/config/conversion/gemma4/gemma-4-12b-it-text-q4k-ehf16-af16.json +31 -5
- package/src/config/conversion/gemma4/gemma-4-12b-it-text-q4k-ehf16-af32.json +30 -4
- package/src/config/conversion/gemma4/gemma-4-12b-it-text-q4k-ehf16-hq4k-af16.json +30 -4
- package/src/config/conversion/gemma4/gemma-4-12b-it-text-w4a16-ct-ehf16-af16.json +30 -4
- package/src/config/conversion/gemma4/gemma-4-31b-it-text-q4k-ehf16-af16.json +30 -5
- package/src/config/conversion/gemma4/gemma-4-31b-it-text-q4k-ehf16-af32.json +29 -4
- package/src/config/conversion/gemma4/gemma-4-e2b-it-q4k-ehf16-af16-int4ple.json +30 -5
- package/src/config/conversion/gemma4/gemma-4-e2b-it-q4k-ehf16-af32-int4ple.json +29 -4
- package/src/config/conversion/gemma4/gemma-4-e2b-it-q4k-ehf16-af32.json +29 -4
- package/src/config/conversion/gemma4/gemma-4-moe-q4k-ehf16-af32.json +29 -4
- package/src/config/conversion/gpt-oss-20b-f16-xmxfp4.json +29 -4
- package/src/config/conversion/janus/janus-pro-1b-text-q4k-ehaf16.json +29 -4
- package/src/config/conversion/lfm2/lfm2.5-1.2b-instruct-q4k-ehf16-af32.json +29 -4
- package/src/config/conversion/minicpm/minicpm4-0-5b-f16-af32.json +481 -0
- package/src/config/conversion/qwen3/qwen-3-5-0-8b-q4k-ehaf16.json +29 -4
- package/src/config/conversion/qwen3/qwen-3-5-2b-q4k-ehaf16.json +29 -4
- package/src/config/conversion/qwen3/qwen-3-6-27b-q4k-eaf16.json +30 -5
- package/src/config/conversion/qwen3/qwen-3-6-27b-q4k-ehaf16.json +29 -4
- package/src/config/conversion/qwen3/qwen-3-embedding-0-6b-q4k-ehf16-af32.json +395 -0
- package/src/config/conversion/qwen3/qwen-3-reranker-0-6b-f16-af32.json +429 -0
- package/src/config/conversion/qwen3/qwen-3-reranker-0-6b-q4k-ehf16-af32.json +548 -0
- package/src/config/execution-contract-check.js +14 -10
- 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 +41 -0
- package/src/config/param-validator.js +67 -8
- package/src/config/required-inference-fields-contract-check.js +14 -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/debug.schema.d.ts +2 -0
- package/src/config/schema/debug.schema.js +1 -0
- package/src/config/schema/doppler.schema.d.ts +5 -1
- package/src/config/schema/doppler.schema.js +117 -292
- 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/inference.schema.d.ts +17 -3
- package/src/config/schema/kvcache.schema.d.ts +19 -3
- package/src/config/schema/kvcache.schema.js +2 -1
- package/src/config/schema/manifest.schema.d.ts +38 -1
- package/src/config/schema/manifest.schema.js +18 -0
- package/src/config/schema/speculation-self.schema.d.ts +16 -0
- package/src/config/source-packages/litert/gemma-4-12b-it.json +20 -1
- package/src/config/source-packages/litert/gemma-4-e2b-it.json +20 -1
- package/src/config/source-packages/litert/gemma-4-e4b-it.json +21 -1
- package/src/config/transforms/capability-transform-resolver.js +7 -1
- package/src/config/transforms/execution-graph-transforms.js +10 -2
- package/src/converter/conversion-plan.js +2 -0
- package/src/converter/parsers/transformer.js +27 -1
- package/src/converter/quantization-info.js +16 -4
- package/src/converter/rope-config.js +55 -3
- package/src/debug/config.d.ts +4 -0
- package/src/experimental/orchestration/index.d.ts +10 -1
- package/src/experimental/training/distillation/student-fixture.js +1 -1
- package/src/experimental/training/lora-pipeline.js +1 -1
- package/src/experimental/training/lora.js +2 -2
- package/src/experimental/training/suite.js +2 -2
- 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/backward/attention_backward.js +8 -0
- package/src/gpu/kernels/dequant.d.ts +6 -6
- package/src/gpu/kernels/dequant.js +17 -6
- 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 +414 -21
- 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 +9 -2
- package/src/gpu/kernels/matmul.d.ts +1 -1
- package/src/gpu/kernels/matmul.js +24 -6
- package/src/gpu/kernels/moe.d.ts +5 -1
- package/src/gpu/kernels/moe.js +5 -2
- 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/softmax.d.ts +3 -3
- package/src/gpu/kernels/softmax.js +9 -2
- 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/kernels/turboquant-codebook.js +1 -1
- package/src/gpu/kernels/types.d.ts +1 -1
- package/src/gpu/profiler.js +11 -1
- 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-model-helpers.js +4 -1
- package/src/inference/browser-harness-suite-helpers.js +1 -1
- package/src/inference/browser-harness-text-helpers.js +589 -117
- package/src/inference/browser-harness.d.ts +3 -3
- package/src/inference/browser-harness.js +486 -13
- package/src/inference/fixtures/rerank-semantic-fixtures.json +38 -0
- package/src/inference/kv-cache/base.js +42 -1
- package/src/inference/kv-cache/quantized.js +16 -2
- package/src/inference/kv-cache/tiered.js +50 -15
- package/src/inference/moe-router.js +38 -7
- package/src/inference/pipelines/diffusion/text-encoder-gpu.js +8 -6
- 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 +23 -1
- package/src/inference/pipelines/text/config.js +235 -112
- package/src/inference/pipelines/text/embed.js +39 -12
- package/src/inference/pipelines/text/execution-plan.d.ts +2 -0
- package/src/inference/pipelines/text/execution-plan.js +45 -2
- 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 +254 -25
- package/src/inference/pipelines/text/ffn/moe.js +4 -3
- package/src/inference/pipelines/text/ffn/standard.js +145 -39
- package/src/inference/pipelines/text/generator-decode-policy.js +1 -0
- package/src/inference/pipelines/text/generator-helpers.js +4 -1
- package/src/inference/pipelines/text/generator-runtime.js +54 -20
- package/src/inference/pipelines/text/generator-steps.js +5 -2
- package/src/inference/pipelines/text/generator.d.ts +18 -1
- package/src/inference/pipelines/text/generator.js +303 -20
- package/src/inference/pipelines/text/init-qkv-fusion.js +155 -62
- package/src/inference/pipelines/text/init-rope.js +54 -6
- package/src/inference/pipelines/text/init.d.ts +22 -3
- package/src/inference/pipelines/text/init.js +65 -17
- package/src/inference/pipelines/text/layer.js +111 -5
- 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 +185 -24
- package/src/inference/pipelines/text/logits/gpu.d.ts +1 -1
- package/src/inference/pipelines/text/logits/gpu.js +78 -115
- package/src/inference/pipelines/text/logits/index.js +105 -87
- package/src/inference/pipelines/text/logits/precision-policy.js +103 -0
- package/src/inference/pipelines/text/logits/types.d.ts +1 -0
- package/src/inference/pipelines/text/logits/utils.d.ts +2 -0
- package/src/inference/pipelines/text/logits/utils.js +11 -0
- package/src/inference/pipelines/text/lora-apply.js +6 -4
- package/src/inference/pipelines/text/model-load.js +6 -0
- package/src/inference/pipelines/text/moe-cpu.js +25 -0
- package/src/inference/pipelines/text/moe-gpu.js +148 -97
- package/src/inference/pipelines/text/moe-shape-validator.js +149 -31
- package/src/inference/pipelines/text/ops.js +48 -1
- package/src/inference/pipelines/text/per-layer-inputs.js +105 -21
- 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/weights.js +7 -2
- package/src/inference/pipelines/text.d.ts +21 -1
- package/src/inference/pipelines/text.js +60 -14
- package/src/inference/tokenizers/behavior-flags.js +23 -2
- 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/final-weights-loader.js +12 -1
- 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/execution-rules-contract-check.js +12 -3
- package/src/rules/inference/capability-transforms.rules.json +159 -22
- package/src/rules/inference/config.rules.json +26 -2
- package/src/rules/inference/dtype.rules.json +12 -5
- package/src/rules/inference/execution.rules.json +5 -19
- package/src/rules/inference/ffn.rules.json +33 -0
- package/src/rules/inference/moe.rules.json +97 -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/command-api-constants.js +2 -2
- package/src/tooling/command-api-family-normalizers.js +2 -1
- package/src/tooling/command-api.d.ts +1 -1
- package/src/tooling/command-runner-shared.js +3 -1
- package/src/tooling/node-browser-command-runner.js +42 -7
- package/src/tooling/node-converter.js +20 -0
- package/src/tooling/node-source-runtime.js +5 -1
- 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-readme-hero.svg +0 -84
- package/assets/doppler-runtime-map.svg +0 -84
- 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 -35
- 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 -58
- 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 -318
- 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 -28
- 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 -26
- 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 -300
- 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 -159
- 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 -124
- 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 -68
- 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 -25
- 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 -89
- 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/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 -40
- 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 -219
- 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
|
@@ -1,428 +0,0 @@
|
|
|
1
|
-
--- src/gpu/kernels/fused_ffn.wgsl
|
|
2
|
-
+++ src/gpu/kernels/fused_ffn_f16.wgsl
|
|
3
|
-
@@ -1,7 +1,8 @@
|
|
4
|
-
-// Fused FFN Kernel (Tier 2 P0)
|
|
5
|
-
+// Fused FFN Kernel - Native F16 I/O
|
|
6
|
-
//
|
|
7
|
-
-// Fuses gate + up weight projections into a single kernel to reduce
|
|
8
|
-
-// memory bandwidth and kernel launch overhead.
|
|
9
|
-
+// Full f16 pipeline variant: f16 input, f16 weights, f16 output.
|
|
10
|
-
+// f32 accumulators and shared memory for numerical stability.
|
|
11
|
-
+// Requires 'shader-f16' feature.
|
|
12
|
-
//
|
|
13
|
-
// Standard FFN (SwiGLU/GeGLU):
|
|
14
|
-
// gate = x @ W_gate
|
|
15
|
-
@@ -9,62 +10,42 @@
|
|
16
|
-
// hidden = activation(gate) * up
|
|
17
|
-
// out = hidden @ W_down
|
|
18
|
-
//
|
|
19
|
-
-// This kernel fuses steps 1-3 into a single kernel:
|
|
20
|
-
-// 1. Load x from global memory (1x)
|
|
21
|
-
-// 2. Compute gate and up projections simultaneously
|
|
22
|
-
-// 3. Apply activation and multiply
|
|
23
|
-
-// 4. Store result
|
|
24
|
-
-//
|
|
25
|
-
-// Memory savings: 2x reduction in x reads, eliminates intermediate buffers
|
|
26
|
-
-//
|
|
27
|
-
-// For Q4_K weights, this kernel:
|
|
28
|
-
-// - Dequantizes gate and up weights on-the-fly
|
|
29
|
-
-// - Uses subgroup operations for reduction
|
|
30
|
-
-// - Achieves 2-3x speedup over separate kernel approach
|
|
31
|
-
-
|
|
32
|
-
+// This kernel fuses steps 1-3 into a single dispatch.
|
|
33
|
-
+// Input is widened f16→f32 on shared memory load; output is narrowed f32→f16 on write.
|
|
34
|
-
+
|
|
35
|
-
+enable f16;
|
|
36
|
-
enable subgroups;
|
|
37
|
-
-
|
|
38
|
-
-// Q4_K constants
|
|
39
|
-
-const QK_K: u32 = 256u; // Elements per super-block
|
|
40
|
-
-const BLOCK_SIZE: u32 = 144u; // Bytes per Q4_K block
|
|
41
|
-
-const SUBBLOCK_SIZE: u32 = 32u; // Elements per sub-block
|
|
42
|
-
|
|
43
|
-
override WORKGROUP_SIZE: u32 = 256u;
|
|
44
|
-
override SHARED_INPUT_SIZE: u32 = 256u;
|
|
45
|
-
-const MAX_SUBGROUPS: u32 = 256u; // Supports subgroup_size >= 1
|
|
46
|
-
-const MAX_SUBGROUPS_PER_OUTPUT: u32 = 64u; // THREADS_PER_OUTPUT max
|
|
47
|
-
+const MAX_SUBGROUPS: u32 = 256u;
|
|
48
|
-
|
|
49
|
-
struct Uniforms {
|
|
50
|
-
- M: u32, // Batch size (usually 1 for decode)
|
|
51
|
-
- hidden_size: u32, // Input dimension
|
|
52
|
-
- intermediate_size: u32, // Output dimension (per gate/up)
|
|
53
|
-
- alpha: f32, // Scale factor
|
|
54
|
-
- activation: u32, // 0=silu, 1=gelu
|
|
55
|
-
- clamp_max: f32, // SwiGLU clamp (0 = disabled)
|
|
56
|
-
- _pad0: u32, // 16-byte alignment padding
|
|
57
|
-
+ M: u32,
|
|
58
|
-
+ hidden_size: u32,
|
|
59
|
-
+ intermediate_size: u32,
|
|
60
|
-
+ alpha: f32,
|
|
61
|
-
+ activation: u32,
|
|
62
|
-
+ clamp_max: f32,
|
|
63
|
-
+ _pad0: u32,
|
|
64
|
-
_pad1: u32,
|
|
65
|
-
}
|
|
66
|
-
|
|
67
|
-
@group(0) @binding(0) var<uniform> u: Uniforms;
|
|
68
|
-
-@group(0) @binding(1) var<storage, read> input: array<f32>;
|
|
69
|
-
-@group(0) @binding(2) var<storage, read> W_gate: array<f32>; // [intermediate_size, hidden_size]
|
|
70
|
-
-@group(0) @binding(3) var<storage, read> W_up: array<f32>; // [intermediate_size, hidden_size]
|
|
71
|
-
-@group(0) @binding(4) var<storage, read_write> output: array<f32>;
|
|
72
|
-
-
|
|
73
|
-
-// Shared memory for input vector (reused for gate and up)
|
|
74
|
-
+@group(0) @binding(1) var<storage, read> input: array<f16>;
|
|
75
|
-
+@group(0) @binding(2) var<storage, read> W_gate: array<f16>;
|
|
76
|
-
+@group(0) @binding(3) var<storage, read> W_up: array<f16>;
|
|
77
|
-
+@group(0) @binding(4) var<storage, read_write> output: array<f16>;
|
|
78
|
-
+
|
|
79
|
-
var<workgroup> shared_input: array<f32, SHARED_INPUT_SIZE>;
|
|
80
|
-
-
|
|
81
|
-
-// For subgroup reduction
|
|
82
|
-
var<workgroup> sg_sums: array<f32, MAX_SUBGROUPS * 2u>;
|
|
83
|
-
|
|
84
|
-
-// SiLU activation: x * sigmoid(x)
|
|
85
|
-
fn silu(x: f32) -> f32 {
|
|
86
|
-
return x / (1.0 + exp(-x));
|
|
87
|
-
}
|
|
88
|
-
|
|
89
|
-
-// GELU activation (approximate)
|
|
90
|
-
fn gelu(x: f32) -> f32 {
|
|
91
|
-
- let c = 0.7978845608; // sqrt(2/pi)
|
|
92
|
-
+ let c = 0.7978845608;
|
|
93
|
-
return 0.5 * x * (1.0 + tanh(c * (x + 0.044715 * x * x * x)));
|
|
94
|
-
}
|
|
95
|
-
|
|
96
|
-
@@ -75,8 +56,7 @@
|
|
97
|
-
return clamp(x, -u.clamp_max, u.clamp_max);
|
|
98
|
-
}
|
|
99
|
-
|
|
100
|
-
-// Fused FFN forward for F32 weights
|
|
101
|
-
-// One workgroup computes one output element
|
|
102
|
-
+// Single-output-per-workgroup variant for decode (M=1)
|
|
103
|
-
@compute @workgroup_size(WORKGROUP_SIZE, 1, 1)
|
|
104
|
-
fn main(
|
|
105
|
-
@builtin(local_invocation_id) lid: vec3<u32>,
|
|
106
|
-
@@ -87,9 +67,8 @@
|
|
107
|
-
let out_idx = wg_id.x;
|
|
108
|
-
let tid = lid.x;
|
|
109
|
-
let hidden_size = u.hidden_size;
|
|
110
|
-
- let intermediate_size = u.intermediate_size;
|
|
111
|
-
-
|
|
112
|
-
- if (out_idx >= intermediate_size) {
|
|
113
|
-
+
|
|
114
|
-
+ if (out_idx >= u.intermediate_size) {
|
|
115
|
-
return;
|
|
116
|
-
}
|
|
117
|
-
|
|
118
|
-
@@ -99,35 +78,31 @@
|
|
119
|
-
var gate_sum: f32 = 0.0;
|
|
120
|
-
var up_sum: f32 = 0.0;
|
|
121
|
-
|
|
122
|
-
- // Weight offsets for this output element
|
|
123
|
-
- let gate_base = out_idx * hidden_size;
|
|
124
|
-
- let up_base = out_idx * hidden_size;
|
|
125
|
-
+ let weight_base = out_idx * hidden_size;
|
|
126
|
-
|
|
127
|
-
let num_tiles = (hidden_size + SHARED_INPUT_SIZE - 1u) / SHARED_INPUT_SIZE;
|
|
128
|
-
for (var tile = 0u; tile < num_tiles; tile = tile + 1u) {
|
|
129
|
-
let tile_start = tile * SHARED_INPUT_SIZE;
|
|
130
|
-
let tile_end = min(tile_start + SHARED_INPUT_SIZE, hidden_size);
|
|
131
|
-
|
|
132
|
-
- // Load input tile into shared memory
|
|
133
|
-
+ // Load f16 input, widen to f32 in shared memory for accumulation precision
|
|
134
|
-
for (var i = tid; i < SHARED_INPUT_SIZE; i = i + WORKGROUP_SIZE) {
|
|
135
|
-
let idx = tile_start + i;
|
|
136
|
-
if (idx < hidden_size) {
|
|
137
|
-
- shared_input[i] = input[idx];
|
|
138
|
-
+ shared_input[i] = f32(input[idx]);
|
|
139
|
-
}
|
|
140
|
-
}
|
|
141
|
-
workgroupBarrier();
|
|
142
|
-
|
|
143
|
-
- // Compute partial sums for this tile
|
|
144
|
-
for (var k = tile_start + tid; k < tile_end; k = k + WORKGROUP_SIZE) {
|
|
145
|
-
let x = shared_input[k - tile_start];
|
|
146
|
-
- gate_sum += x * W_gate[gate_base + k];
|
|
147
|
-
- up_sum += x * W_up[up_base + k];
|
|
148
|
-
- }
|
|
149
|
-
-
|
|
150
|
-
- workgroupBarrier();
|
|
151
|
-
- }
|
|
152
|
-
-
|
|
153
|
-
- // Phase 3: Reduce across threads using subgroups
|
|
154
|
-
+ gate_sum += x * f32(W_gate[weight_base + k]);
|
|
155
|
-
+ up_sum += x * f32(W_up[weight_base + k]);
|
|
156
|
-
+ }
|
|
157
|
-
+
|
|
158
|
-
+ workgroupBarrier();
|
|
159
|
-
+ }
|
|
160
|
-
+
|
|
161
|
-
let sg_gate = subgroupAdd(gate_sum);
|
|
162
|
-
let sg_up = subgroupAdd(up_sum);
|
|
163
|
-
|
|
164
|
-
@@ -137,7 +112,6 @@
|
|
165
|
-
}
|
|
166
|
-
workgroupBarrier();
|
|
167
|
-
|
|
168
|
-
- // Thread 0 does final reduction
|
|
169
|
-
if (tid == 0u) {
|
|
170
|
-
var final_gate: f32 = 0.0;
|
|
171
|
-
var final_up: f32 = 0.0;
|
|
172
|
-
@@ -146,7 +120,6 @@
|
|
173
|
-
final_up += sg_sums[s + MAX_SUBGROUPS];
|
|
174
|
-
}
|
|
175
|
-
|
|
176
|
-
- // Apply activation and multiply
|
|
177
|
-
var activated: f32;
|
|
178
|
-
if (u.activation == 0u) {
|
|
179
|
-
activated = silu(final_gate);
|
|
180
|
-
@@ -154,191 +127,11 @@
|
|
181
|
-
activated = gelu(final_gate);
|
|
182
|
-
}
|
|
183
|
-
|
|
184
|
-
- output[out_idx] = clamp_swiglu(activated * final_up * u.alpha);
|
|
185
|
-
- }
|
|
186
|
-
-}
|
|
187
|
-
-
|
|
188
|
-
-// Optimized variant: Multiple outputs per workgroup
|
|
189
|
-
-// For better GPU utilization when intermediate_size is small
|
|
190
|
-
-const OUTPUTS_PER_WG: u32 = 4u;
|
|
191
|
-
-const THREADS_PER_OUTPUT: u32 = 64u;
|
|
192
|
-
-
|
|
193
|
-
-var<workgroup> multi_sg_sums: array<f32, OUTPUTS_PER_WG * MAX_SUBGROUPS_PER_OUTPUT * 2u>;
|
|
194
|
-
-
|
|
195
|
-
-@compute @workgroup_size(WORKGROUP_SIZE, 1, 1)
|
|
196
|
-
-fn main_multi(
|
|
197
|
-
- @builtin(local_invocation_id) lid: vec3<u32>,
|
|
198
|
-
- @builtin(workgroup_id) wg_id: vec3<u32>,
|
|
199
|
-
- @builtin(subgroup_invocation_id) sg_id: u32,
|
|
200
|
-
- @builtin(subgroup_size) sg_size: u32,
|
|
201
|
-
-) {
|
|
202
|
-
- let tid = lid.x;
|
|
203
|
-
- let out_in_wg = tid / THREADS_PER_OUTPUT;
|
|
204
|
-
- let tid_in_out = tid % THREADS_PER_OUTPUT;
|
|
205
|
-
- let out_idx = wg_id.x * OUTPUTS_PER_WG + out_in_wg;
|
|
206
|
-
-
|
|
207
|
-
- let hidden_size = u.hidden_size;
|
|
208
|
-
- let intermediate_size = u.intermediate_size;
|
|
209
|
-
- let is_active = out_idx < intermediate_size;
|
|
210
|
-
-
|
|
211
|
-
- var gate_sum: f32 = 0.0;
|
|
212
|
-
- var up_sum: f32 = 0.0;
|
|
213
|
-
-
|
|
214
|
-
- let gate_base = out_idx * hidden_size;
|
|
215
|
-
- let up_base = out_idx * hidden_size;
|
|
216
|
-
-
|
|
217
|
-
- let num_tiles = (hidden_size + SHARED_INPUT_SIZE - 1u) / SHARED_INPUT_SIZE;
|
|
218
|
-
- for (var tile = 0u; tile < num_tiles; tile = tile + 1u) {
|
|
219
|
-
- let tile_start = tile * SHARED_INPUT_SIZE;
|
|
220
|
-
- let tile_end = min(tile_start + SHARED_INPUT_SIZE, hidden_size);
|
|
221
|
-
-
|
|
222
|
-
- // Load input tile
|
|
223
|
-
- if (tid < SHARED_INPUT_SIZE) {
|
|
224
|
-
- let idx = tile_start + tid;
|
|
225
|
-
- if (idx < hidden_size) {
|
|
226
|
-
- shared_input[tid] = input[idx];
|
|
227
|
-
- }
|
|
228
|
-
- }
|
|
229
|
-
- workgroupBarrier();
|
|
230
|
-
-
|
|
231
|
-
- // Strided access pattern per output
|
|
232
|
-
- if (is_active) {
|
|
233
|
-
- for (var k = tile_start + tid_in_out; k < tile_end; k = k + THREADS_PER_OUTPUT) {
|
|
234
|
-
- let x = shared_input[k - tile_start];
|
|
235
|
-
- gate_sum += x * W_gate[gate_base + k];
|
|
236
|
-
- up_sum += x * W_up[up_base + k];
|
|
237
|
-
- }
|
|
238
|
-
- }
|
|
239
|
-
- workgroupBarrier();
|
|
240
|
-
- }
|
|
241
|
-
-
|
|
242
|
-
- // Reduce within each output group
|
|
243
|
-
- let local_sg_id = tid_in_out / sg_size;
|
|
244
|
-
- let sg_gate = subgroupAdd(gate_sum);
|
|
245
|
-
- let sg_up = subgroupAdd(up_sum);
|
|
246
|
-
-
|
|
247
|
-
- if (is_active && sg_id == 0u && local_sg_id < MAX_SUBGROUPS_PER_OUTPUT) {
|
|
248
|
-
- let base = out_in_wg * MAX_SUBGROUPS_PER_OUTPUT * 2u;
|
|
249
|
-
- multi_sg_sums[base + local_sg_id] = sg_gate;
|
|
250
|
-
- multi_sg_sums[base + MAX_SUBGROUPS_PER_OUTPUT + local_sg_id] = sg_up;
|
|
251
|
-
- }
|
|
252
|
-
- workgroupBarrier();
|
|
253
|
-
-
|
|
254
|
-
- // First thread of each output finalizes
|
|
255
|
-
- if (is_active && tid_in_out == 0u) {
|
|
256
|
-
- let num_sgs = min(MAX_SUBGROUPS_PER_OUTPUT, (THREADS_PER_OUTPUT + sg_size - 1u) / sg_size);
|
|
257
|
-
- var final_gate: f32 = 0.0;
|
|
258
|
-
- var final_up: f32 = 0.0;
|
|
259
|
-
-
|
|
260
|
-
- let base = out_in_wg * MAX_SUBGROUPS_PER_OUTPUT * 2u;
|
|
261
|
-
- for (var s = 0u; s < num_sgs; s++) {
|
|
262
|
-
- final_gate += multi_sg_sums[base + s];
|
|
263
|
-
- final_up += multi_sg_sums[base + MAX_SUBGROUPS_PER_OUTPUT + s];
|
|
264
|
-
- }
|
|
265
|
-
-
|
|
266
|
-
- var activated: f32;
|
|
267
|
-
- if (u.activation == 0u) {
|
|
268
|
-
- activated = silu(final_gate);
|
|
269
|
-
- } else {
|
|
270
|
-
- activated = gelu(final_gate);
|
|
271
|
-
- }
|
|
272
|
-
-
|
|
273
|
-
- output[out_idx] = clamp_swiglu(activated * final_up * u.alpha);
|
|
274
|
-
- }
|
|
275
|
-
-}
|
|
276
|
-
-
|
|
277
|
-
-// F16 weights variant - optimized for memory bandwidth
|
|
278
|
-
-@compute @workgroup_size(WORKGROUP_SIZE, 1, 1)
|
|
279
|
-
-fn main_f16(
|
|
280
|
-
- @builtin(local_invocation_id) lid: vec3<u32>,
|
|
281
|
-
- @builtin(workgroup_id) wg_id: vec3<u32>,
|
|
282
|
-
- @builtin(subgroup_invocation_id) sg_id: u32,
|
|
283
|
-
- @builtin(subgroup_size) sg_size: u32,
|
|
284
|
-
-) {
|
|
285
|
-
- let out_idx = wg_id.x;
|
|
286
|
-
- let tid = lid.x;
|
|
287
|
-
- let hidden_size = u.hidden_size;
|
|
288
|
-
- let intermediate_size = u.intermediate_size;
|
|
289
|
-
-
|
|
290
|
-
- if (out_idx >= intermediate_size) {
|
|
291
|
-
- return;
|
|
292
|
-
- }
|
|
293
|
-
-
|
|
294
|
-
- let subgroup_id = tid / sg_size;
|
|
295
|
-
- let num_subgroups = (WORKGROUP_SIZE + sg_size - 1u) / sg_size;
|
|
296
|
-
-
|
|
297
|
-
- var gate_sum: f32 = 0.0;
|
|
298
|
-
- var up_sum: f32 = 0.0;
|
|
299
|
-
-
|
|
300
|
-
- // Each thread handles stride of WORKGROUP_SIZE
|
|
301
|
-
- let gate_base = out_idx * hidden_size;
|
|
302
|
-
- let up_base = out_idx * hidden_size;
|
|
303
|
-
-
|
|
304
|
-
- let num_tiles = (hidden_size + SHARED_INPUT_SIZE - 1u) / SHARED_INPUT_SIZE;
|
|
305
|
-
- for (var tile = 0u; tile < num_tiles; tile = tile + 1u) {
|
|
306
|
-
- let tile_start = tile * SHARED_INPUT_SIZE;
|
|
307
|
-
- let tile_end = min(tile_start + SHARED_INPUT_SIZE, hidden_size);
|
|
308
|
-
-
|
|
309
|
-
- // Load input tile
|
|
310
|
-
- for (var i = tid; i < SHARED_INPUT_SIZE; i = i + WORKGROUP_SIZE) {
|
|
311
|
-
- let idx = tile_start + i;
|
|
312
|
-
- if (idx < hidden_size) {
|
|
313
|
-
- shared_input[i] = input[idx];
|
|
314
|
-
- }
|
|
315
|
-
- }
|
|
316
|
-
- workgroupBarrier();
|
|
317
|
-
-
|
|
318
|
-
- for (var k = tile_start + tid; k < tile_end; k = k + WORKGROUP_SIZE) {
|
|
319
|
-
- let x = shared_input[k - tile_start];
|
|
320
|
-
-
|
|
321
|
-
- // Read F16 weights (packed as u32)
|
|
322
|
-
- let gate_packed = W_gate[gate_base / 2u + k / 2u];
|
|
323
|
-
- let up_packed = W_up[up_base / 2u + k / 2u];
|
|
324
|
-
-
|
|
325
|
-
- let gate_vec = unpack2x16float(bitcast<u32>(gate_packed));
|
|
326
|
-
- let up_vec = unpack2x16float(bitcast<u32>(up_packed));
|
|
327
|
-
-
|
|
328
|
-
- let g = select(gate_vec.y, gate_vec.x, (k % 2u) == 0u);
|
|
329
|
-
- let u = select(up_vec.y, up_vec.x, (k % 2u) == 0u);
|
|
330
|
-
-
|
|
331
|
-
- gate_sum += x * g;
|
|
332
|
-
- up_sum += x * u;
|
|
333
|
-
- }
|
|
334
|
-
-
|
|
335
|
-
- workgroupBarrier();
|
|
336
|
-
- }
|
|
337
|
-
-
|
|
338
|
-
- // Reduce
|
|
339
|
-
- let sg_gate = subgroupAdd(gate_sum);
|
|
340
|
-
- let sg_up = subgroupAdd(up_sum);
|
|
341
|
-
-
|
|
342
|
-
- if (sg_id == 0u && subgroup_id < num_subgroups) {
|
|
343
|
-
- sg_sums[subgroup_id] = sg_gate;
|
|
344
|
-
- sg_sums[subgroup_id + MAX_SUBGROUPS] = sg_up;
|
|
345
|
-
- }
|
|
346
|
-
- workgroupBarrier();
|
|
347
|
-
-
|
|
348
|
-
- if (tid == 0u) {
|
|
349
|
-
- var final_gate: f32 = 0.0;
|
|
350
|
-
- var final_up: f32 = 0.0;
|
|
351
|
-
- for (var s = 0u; s < num_subgroups; s++) {
|
|
352
|
-
- final_gate += sg_sums[s];
|
|
353
|
-
- final_up += sg_sums[s + MAX_SUBGROUPS];
|
|
354
|
-
- }
|
|
355
|
-
-
|
|
356
|
-
- var activated: f32;
|
|
357
|
-
- if (u.activation == 0u) {
|
|
358
|
-
- activated = silu(final_gate);
|
|
359
|
-
- } else {
|
|
360
|
-
- activated = gelu(final_gate);
|
|
361
|
-
- }
|
|
362
|
-
-
|
|
363
|
-
- output[out_idx] = clamp_swiglu(activated * final_up * u.alpha);
|
|
364
|
-
+ output[out_idx] = f16(clamp_swiglu(activated * final_up * u.alpha));
|
|
365
|
-
}
|
|
366
|
-
}
|
|
367
|
-
|
|
368
|
-
// Batched variant for prefill (M > 1)
|
|
369
|
-
-// Each workgroup handles one output element across all batch items
|
|
370
|
-
@compute @workgroup_size(WORKGROUP_SIZE, 1, 1)
|
|
371
|
-
fn main_batched(
|
|
372
|
-
@builtin(local_invocation_id) lid: vec3<u32>,
|
|
373
|
-
@@ -351,9 +144,8 @@
|
|
374
|
-
let tid = lid.x;
|
|
375
|
-
let hidden_size = u.hidden_size;
|
|
376
|
-
let intermediate_size = u.intermediate_size;
|
|
377
|
-
- let M = u.M;
|
|
378
|
-
-
|
|
379
|
-
- if (out_idx >= intermediate_size || batch_idx >= M) {
|
|
380
|
-
+
|
|
381
|
-
+ if (out_idx >= intermediate_size || batch_idx >= u.M) {
|
|
382
|
-
return;
|
|
383
|
-
}
|
|
384
|
-
|
|
385
|
-
@@ -363,8 +155,7 @@
|
|
386
|
-
var gate_sum: f32 = 0.0;
|
|
387
|
-
var up_sum: f32 = 0.0;
|
|
388
|
-
|
|
389
|
-
- let gate_base = out_idx * hidden_size;
|
|
390
|
-
- let up_base = out_idx * hidden_size;
|
|
391
|
-
+ let weight_base = out_idx * hidden_size;
|
|
392
|
-
let input_base = batch_idx * hidden_size;
|
|
393
|
-
|
|
394
|
-
let num_tiles = (hidden_size + SHARED_INPUT_SIZE - 1u) / SHARED_INPUT_SIZE;
|
|
395
|
-
@@ -372,19 +163,19 @@
|
|
396
|
-
let tile_start = tile * SHARED_INPUT_SIZE;
|
|
397
|
-
let tile_end = min(tile_start + SHARED_INPUT_SIZE, hidden_size);
|
|
398
|
-
|
|
399
|
-
- // Load input tile for this batch item
|
|
400
|
-
+ // Load f16 input, widen to f32 in shared memory for accumulation precision
|
|
401
|
-
for (var i = tid; i < SHARED_INPUT_SIZE; i = i + WORKGROUP_SIZE) {
|
|
402
|
-
let idx = tile_start + i;
|
|
403
|
-
if (idx < hidden_size) {
|
|
404
|
-
- shared_input[i] = input[input_base + idx];
|
|
405
|
-
+ shared_input[i] = f32(input[input_base + idx]);
|
|
406
|
-
}
|
|
407
|
-
}
|
|
408
|
-
workgroupBarrier();
|
|
409
|
-
|
|
410
|
-
for (var k = tile_start + tid; k < tile_end; k = k + WORKGROUP_SIZE) {
|
|
411
|
-
let x = shared_input[k - tile_start];
|
|
412
|
-
- gate_sum += x * W_gate[gate_base + k];
|
|
413
|
-
- up_sum += x * W_up[up_base + k];
|
|
414
|
-
+ gate_sum += x * f32(W_gate[weight_base + k]);
|
|
415
|
-
+ up_sum += x * f32(W_up[weight_base + k]);
|
|
416
|
-
}
|
|
417
|
-
|
|
418
|
-
workgroupBarrier();
|
|
419
|
-
@@ -415,6 +206,6 @@
|
|
420
|
-
}
|
|
421
|
-
|
|
422
|
-
let out_offset = batch_idx * intermediate_size + out_idx;
|
|
423
|
-
- output[out_offset] = clamp_swiglu(activated * final_up * u.alpha);
|
|
424
|
-
- }
|
|
425
|
-
-}
|
|
426
|
-
+ output[out_offset] = f16(clamp_swiglu(activated * final_up * u.alpha));
|
|
427
|
-
+ }
|
|
428
|
-
+}
|
|
@@ -1,114 +0,0 @@
|
|
|
1
|
-
--- src/gpu/kernels/fused_matmul_q4_batched.wgsl
|
|
2
|
-
+++ src/gpu/kernels/fused_matmul_q4_batched_f16.wgsl
|
|
3
|
-
@@ -1,25 +1,14 @@
|
|
4
|
-
-// Fused Q4_K Matmul Kernel - W4A16
|
|
5
|
-
-// Directly computes C = A * dequant(B_q4k) without separate dequant pass
|
|
6
|
-
+// Fused Q4_K Matmul Kernel - W4A16 (F16 output, batched)
|
|
7
|
-
//
|
|
8
|
-
-// For M=1 decode (GEMV): C[N] = A[K] * B_q4k^T[N,K]
|
|
9
|
-
-// B_q4k is stored in Q4_K format: [N * ceil(K/256) * 144 bytes]
|
|
10
|
-
-//
|
|
11
|
-
-// Key optimizations:
|
|
12
|
-
-// 1. Fused dequant + matmul - eliminates memory round-trip (2-3x speedup)
|
|
13
|
-
-// 2. Subgroup operations for reduction
|
|
14
|
-
-// 3. On-the-fly dequantization in registers
|
|
15
|
-
-//
|
|
16
|
-
-// A is f32 (activations), B_q4k is Q4_K quantized weights, C is f32.
|
|
17
|
-
+// Batched version for prefill (M > 1) with f16 output.
|
|
18
|
-
|
|
19
|
-
+enable f16;
|
|
20
|
-
enable subgroups;
|
|
21
|
-
|
|
22
|
-
// Q4_K constants
|
|
23
|
-
const QK_K: u32 = 256u; // Elements per super-block
|
|
24
|
-
const BLOCK_SIZE: u32 = 144u; // Bytes per Q4_K block
|
|
25
|
-
const SUBBLOCK_SIZE: u32 = 32u; // Elements per sub-block
|
|
26
|
-
-
|
|
27
|
-
-override WORKGROUP_SIZE: u32 = 256u;
|
|
28
|
-
-const MAX_SUBGROUPS: u32 = 256u; // Supports subgroup_size >= 1
|
|
29
|
-
|
|
30
|
-
struct Uniforms {
|
|
31
|
-
M: u32, // Always 1 for GEMV
|
|
32
|
-
@@ -43,10 +32,7 @@
|
|
33
|
-
@group(0) @binding(0) var<uniform> u: Uniforms;
|
|
34
|
-
@group(0) @binding(1) var<storage, read> A: array<f32>;
|
|
35
|
-
@group(0) @binding(2) var<storage, read> B_q4k: array<Q4KBlock>;
|
|
36
|
-
-@group(0) @binding(3) var<storage, read_write> C: array<f32>;
|
|
37
|
-
-
|
|
38
|
-
-// Shared memory for subgroup reduction
|
|
39
|
-
-var<workgroup> wg_sums: array<f32, MAX_SUBGROUPS>;
|
|
40
|
-
+@group(0) @binding(4) var<storage, read_write> C_f16: array<f16>;
|
|
41
|
-
|
|
42
|
-
// Extract f16 from packed u32
|
|
43
|
-
fn unpack_f16_lo(packed: u32) -> f32 {
|
|
44
|
-
@@ -105,10 +91,8 @@
|
|
45
|
-
// Batched version for prefill (M > 1)
|
|
46
|
-
// Uses 2D dispatch: workgroup (x,y) computes output C[y*TILE_M : (y+1)*TILE_M, x]
|
|
47
|
-
// Each workgroup computes TILE_M rows × 1 column
|
|
48
|
-
-//
|
|
49
|
-
-// IMPORTANT: Previous version used 16 threads per column which caused subgroup
|
|
50
|
-
-// mixing when sg_size=32 (two columns in same subgroup). This version uses
|
|
51
|
-
-// 64 threads per column to ensure correct subgroup reduction.
|
|
52
|
-
+// ============================================================================
|
|
53
|
-
+
|
|
54
|
-
const MAX_TILE_M: u32 = 4u;
|
|
55
|
-
const MAX_THREADS_PER_COL: u32 = 64u;
|
|
56
|
-
const MAX_SUBGROUPS_PER_ROW: u32 = 64u; // Support sg_size >= 1 (64/1 = 64)
|
|
57
|
-
@@ -120,7 +104,7 @@
|
|
58
|
-
var<workgroup> batched_wg_sums: array<f32, MAX_TILE_M * MAX_SUBGROUPS_PER_ROW>;
|
|
59
|
-
|
|
60
|
-
@compute @workgroup_size(THREADS_PER_COL, TILE_M, 1)
|
|
61
|
-
-fn main_batched(
|
|
62
|
-
+fn main_batched_f16(
|
|
63
|
-
@builtin(local_invocation_id) lid: vec3<u32>,
|
|
64
|
-
@builtin(workgroup_id) wg_id: vec3<u32>,
|
|
65
|
-
@builtin(subgroup_invocation_id) sg_id: u32,
|
|
66
|
-
@@ -132,19 +116,13 @@
|
|
67
|
-
|
|
68
|
-
let local_id = lid.x;
|
|
69
|
-
let row = wg_id.y * TILE_M + lid.y;
|
|
70
|
-
- let col = wg_id.x; // One column per workgroup X (no more /16 mixing)
|
|
71
|
-
-
|
|
72
|
-
- // Track validity - NO early return to maintain uniform control flow for subgroupAdd
|
|
73
|
-
+ let col = wg_id.x;
|
|
74
|
-
let is_valid = row < u.M && col < u.N;
|
|
75
|
-
|
|
76
|
-
var partial_sum: f32 = 0.0;
|
|
77
|
-
|
|
78
|
-
- // Only do work if this output cell is valid
|
|
79
|
-
if (is_valid) {
|
|
80
|
-
let num_blocks = u.num_blocks_per_row;
|
|
81
|
-
-
|
|
82
|
-
- // B_q4k layout: row-major [N, K/256] - block b for column col is at col * num_blocks + b
|
|
83
|
-
- // Each thread processes every 64th block (instead of 16th)
|
|
84
|
-
for (var b: u32 = local_id; b < num_blocks; b = b + THREADS_PER_COL) {
|
|
85
|
-
let block = B_q4k[col * num_blocks + b];
|
|
86
|
-
let d = unpack_f16_lo(block.d_dmin);
|
|
87
|
-
@@ -168,12 +146,9 @@
|
|
88
|
-
}
|
|
89
|
-
}
|
|
90
|
-
}
|
|
91
|
-
- } // end if (is_valid)
|
|
92
|
-
+ }
|
|
93
|
-
|
|
94
|
-
- // Subgroup reduction - ALL threads must execute (uniform control flow)
|
|
95
|
-
let sg_sum = subgroupAdd(partial_sum);
|
|
96
|
-
-
|
|
97
|
-
- // Store subgroup results to shared memory (per-row)
|
|
98
|
-
let num_subgroups = (THREADS_PER_COL + sg_size - 1u) / sg_size;
|
|
99
|
-
|
|
100
|
-
if (sg_id == 0u && local_id < THREADS_PER_COL) {
|
|
101
|
-
@@ -183,12 +158,11 @@
|
|
102
|
-
|
|
103
|
-
workgroupBarrier();
|
|
104
|
-
|
|
105
|
-
- // Thread 0 of each row does final reduction and writes result
|
|
106
|
-
if (local_id == 0u && is_valid) {
|
|
107
|
-
var final_sum: f32 = 0.0;
|
|
108
|
-
for (var i: u32 = 0u; i < num_subgroups; i = i + 1u) {
|
|
109
|
-
final_sum = final_sum + batched_wg_sums[lid.y * MAX_SUBGROUPS_PER_ROW + i];
|
|
110
|
-
}
|
|
111
|
-
- C[row * u.N + col] = final_sum * u.alpha;
|
|
112
|
-
+ C_f16[row * u.N + col] = f16(final_sum * u.alpha);
|
|
113
|
-
}
|
|
114
|
-
}
|