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
|
@@ -1,292 +0,0 @@
|
|
|
1
|
-
--- src/gpu/kernels/rope.wgsl
|
|
2
|
-
+++ src/gpu/kernels/rope_f16.wgsl
|
|
3
|
-
@@ -1,19 +1,16 @@
|
|
4
|
-
-// Rotary Position Embeddings (RoPE) Kernel
|
|
5
|
-
+// Rotary Position Embeddings (RoPE) Kernel (F16)
|
|
6
|
-
//
|
|
7
|
-
// Applies rotary position embeddings to Q and K tensors.
|
|
8
|
-
-// RoPE rotates pairs of dimensions based on position and frequency.
|
|
9
|
-
-//
|
|
10
|
-
-// For each pair (x_i, x_{i+1}):
|
|
11
|
-
-// x'_i = x_i * cos(θ) - x_{i+1} * sin(θ)
|
|
12
|
-
-// x'_{i+1} = x_i * sin(θ) + x_{i+1} * cos(θ)
|
|
13
|
-
-//
|
|
14
|
-
-// Where θ = pos * freq_i, freq_i = 1 / (base^(2i/d))
|
|
15
|
-
+// Same math as rope.wgsl, but input/output are f16.
|
|
16
|
-
+// The main entry keeps apply-time arithmetic in f16 for the all-f16 lane.
|
|
17
|
-
//
|
|
18
|
-
// Supports:
|
|
19
|
-
// - Original RoPE (base = 10000)
|
|
20
|
-
// - Scaled RoPE (for extended context)
|
|
21
|
-
// - NTK-aware scaling
|
|
22
|
-
|
|
23
|
-
+enable f16;
|
|
24
|
-
+
|
|
25
|
-
override WORKGROUP_SIZE: u32 = 256u;
|
|
26
|
-
|
|
27
|
-
// YaRN parameters (model-family-specific, config-driven)
|
|
28
|
-
@@ -25,24 +22,24 @@
|
|
29
|
-
const PI: f32 = 3.14159265359;
|
|
30
|
-
|
|
31
|
-
struct Uniforms {
|
|
32
|
-
- seq_len: u32, // Sequence length
|
|
33
|
-
- num_heads: u32, // Number of heads
|
|
34
|
-
- head_dim: u32, // Dimension per head (must be even)
|
|
35
|
-
- start_pos: u32, // Starting position (for decode)
|
|
36
|
-
- rope_base: f32, // Base frequency (default 10000)
|
|
37
|
-
- rope_scale: f32, // Scaling factor for extended context
|
|
38
|
-
- rotary_dim: u32, // Rotary slice within head_dim
|
|
39
|
-
- interleaved: u32, // 1 = adjacent pairs, 0 = rotate-half
|
|
40
|
-
- pair_span_dim: u32, // Rotate-half partner span within head_dim
|
|
41
|
-
+ seq_len: u32,
|
|
42
|
-
+ num_heads: u32,
|
|
43
|
-
+ head_dim: u32,
|
|
44
|
-
+ start_pos: u32,
|
|
45
|
-
+ rope_base: f32,
|
|
46
|
-
+ rope_scale: f32,
|
|
47
|
-
+ rotary_dim: u32,
|
|
48
|
-
+ interleaved: u32,
|
|
49
|
-
+ pair_span_dim: u32,
|
|
50
|
-
_pad0: u32,
|
|
51
|
-
_pad1: u32,
|
|
52
|
-
_pad2: u32,
|
|
53
|
-
}
|
|
54
|
-
|
|
55
|
-
@group(0) @binding(0) var<uniform> u: Uniforms;
|
|
56
|
-
-@group(0) @binding(1) var<storage, read_write> input: array<f32>; // [seq_len, num_heads, head_dim]
|
|
57
|
-
-@group(0) @binding(2) var<storage, read_write> freqs_cos: array<f32>; // Precomputed cos [max_seq_len, head_dim/2]
|
|
58
|
-
-@group(0) @binding(3) var<storage, read_write> freqs_sin: array<f32>; // Precomputed sin [max_seq_len, head_dim/2]
|
|
59
|
-
+@group(0) @binding(1) var<storage, read_write> input: array<f16>;
|
|
60
|
-
+@group(0) @binding(2) var<storage, read_write> freqs_cos: array<f32>;
|
|
61
|
-
+@group(0) @binding(3) var<storage, read_write> freqs_sin: array<f32>;
|
|
62
|
-
|
|
63
|
-
fn get_first_rotary_idx(pair_idx: u32) -> u32 {
|
|
64
|
-
if (u.interleaved == 1u) {
|
|
65
|
-
@@ -68,7 +65,6 @@
|
|
66
|
-
let seq_len = u.seq_len;
|
|
67
|
-
let start_pos = u.start_pos;
|
|
68
|
-
|
|
69
|
-
- // Global thread index (one thread per complex pair)
|
|
70
|
-
let rotary_dim = u.rotary_dim;
|
|
71
|
-
let half_dim = rotary_dim / 2u;
|
|
72
|
-
let total_pairs = seq_len * num_heads * half_dim;
|
|
73
|
-
@@ -78,17 +74,15 @@
|
|
74
|
-
return;
|
|
75
|
-
}
|
|
76
|
-
|
|
77
|
-
- // Decompose index
|
|
78
|
-
let pos = idx / (num_heads * half_dim);
|
|
79
|
-
let remainder = idx % (num_heads * half_dim);
|
|
80
|
-
let head_idx = remainder / half_dim;
|
|
81
|
-
let pair_idx = remainder % half_dim;
|
|
82
|
-
let actual_pos = start_pos + pos;
|
|
83
|
-
|
|
84
|
-
- // Get precomputed cos/sin for this position and dimension
|
|
85
|
-
let freq_idx = actual_pos * half_dim + pair_idx;
|
|
86
|
-
- let cos_val = freqs_cos[freq_idx];
|
|
87
|
-
- let sin_val = freqs_sin[freq_idx];
|
|
88
|
-
+ let cos_val = f16(freqs_cos[freq_idx]);
|
|
89
|
-
+ let sin_val = f16(freqs_sin[freq_idx]);
|
|
90
|
-
|
|
91
|
-
let base_idx = pos * num_heads * head_dim + head_idx * head_dim;
|
|
92
|
-
let first_idx = get_first_rotary_idx(pair_idx);
|
|
93
|
-
@@ -96,11 +90,9 @@
|
|
94
|
-
let x0 = input[base_idx + first_idx];
|
|
95
|
-
let x1 = input[base_idx + second_idx];
|
|
96
|
-
|
|
97
|
-
- // Apply rotation
|
|
98
|
-
let y0 = x0 * cos_val - x1 * sin_val;
|
|
99
|
-
let y1 = x0 * sin_val + x1 * cos_val;
|
|
100
|
-
|
|
101
|
-
- // Write back
|
|
102
|
-
input[base_idx + first_idx] = y0;
|
|
103
|
-
input[base_idx + second_idx] = y1;
|
|
104
|
-
}
|
|
105
|
-
@@ -127,7 +119,6 @@
|
|
106
|
-
return;
|
|
107
|
-
}
|
|
108
|
-
|
|
109
|
-
- // Decompose index
|
|
110
|
-
let pos = idx / (num_heads * half_dim);
|
|
111
|
-
let remainder = idx % (num_heads * half_dim);
|
|
112
|
-
let head_idx = remainder / half_dim;
|
|
113
|
-
@@ -135,7 +126,6 @@
|
|
114
|
-
|
|
115
|
-
let actual_pos = f32(start_pos + pos) / rope_scale;
|
|
116
|
-
|
|
117
|
-
- // Compute frequency: 1 / (base^(2*pair_idx/head_dim))
|
|
118
|
-
let exponent = f32(pair_idx * 2u) / f32(pair_span_dim);
|
|
119
|
-
let freq = 1.0 / pow(rope_base, exponent);
|
|
120
|
-
let theta = actual_pos * freq;
|
|
121
|
-
@@ -146,16 +136,17 @@
|
|
122
|
-
let base_idx = pos * num_heads * head_dim + head_idx * head_dim;
|
|
123
|
-
let first_idx = get_first_rotary_idx(pair_idx);
|
|
124
|
-
let second_idx = get_second_rotary_idx(pair_idx, pair_span_dim);
|
|
125
|
-
- let x0 = input[base_idx + first_idx];
|
|
126
|
-
- let x1 = input[base_idx + second_idx];
|
|
127
|
-
+ let x0 = f32(input[base_idx + first_idx]);
|
|
128
|
-
+ let x1 = f32(input[base_idx + second_idx]);
|
|
129
|
-
+
|
|
130
|
-
+ let y0 = x0 * cos_val - x1 * sin_val;
|
|
131
|
-
+ let y1 = x0 * sin_val + x1 * cos_val;
|
|
132
|
-
|
|
133
|
-
- // Apply rotation
|
|
134
|
-
- input[base_idx + first_idx] = x0 * cos_val - x1 * sin_val;
|
|
135
|
-
- input[base_idx + second_idx] = x0 * sin_val + x1 * cos_val;
|
|
136
|
-
+ input[base_idx + first_idx] = f16(y0);
|
|
137
|
-
+ input[base_idx + second_idx] = f16(y1);
|
|
138
|
-
}
|
|
139
|
-
|
|
140
|
-
// Apply RoPE to both Q and K in one pass
|
|
141
|
-
-// Input layout: Q and K concatenated [seq_len, num_heads, head_dim * 2]
|
|
142
|
-
@compute @workgroup_size(WORKGROUP_SIZE, 1, 1)
|
|
143
|
-
fn rope_qk(
|
|
144
|
-
@builtin(global_invocation_id) global_id: vec3<u32>
|
|
145
|
-
@@ -170,7 +161,6 @@
|
|
146
|
-
let pair_span_dim = u.pair_span_dim;
|
|
147
|
-
|
|
148
|
-
let idx = global_id.x;
|
|
149
|
-
- // Each thread handles one Q-K pair at one dimension pair
|
|
150
|
-
let half_dim = rotary_dim / 2u;
|
|
151
|
-
let total_pairs = seq_len * num_heads * half_dim;
|
|
152
|
-
|
|
153
|
-
@@ -185,7 +175,6 @@
|
|
154
|
-
|
|
155
|
-
let actual_pos = f32(start_pos + pos) / rope_scale;
|
|
156
|
-
|
|
157
|
-
- // Compute frequency
|
|
158
|
-
let exponent = f32(pair_idx * 2u) / f32(pair_span_dim);
|
|
159
|
-
let freq = 1.0 / pow(rope_base, exponent);
|
|
160
|
-
let theta = actual_pos * freq;
|
|
161
|
-
@@ -193,33 +182,28 @@
|
|
162
|
-
let cos_val = cos(theta);
|
|
163
|
-
let sin_val = sin(theta);
|
|
164
|
-
|
|
165
|
-
- // Q is in first half, K in second half
|
|
166
|
-
let q_base_idx = pos * num_heads * head_dim * 2u + head_idx * head_dim;
|
|
167
|
-
- let k_base_idx = q_base_idx + head_dim; // K starts after Q
|
|
168
|
-
+ let k_base_idx = q_base_idx + head_dim;
|
|
169
|
-
|
|
170
|
-
- // Process Q
|
|
171
|
-
let first_idx = get_first_rotary_idx(pair_idx);
|
|
172
|
-
let second_idx = get_second_rotary_idx(pair_idx, pair_span_dim);
|
|
173
|
-
- let q0 = input[q_base_idx + first_idx];
|
|
174
|
-
- let q1 = input[q_base_idx + second_idx];
|
|
175
|
-
- input[q_base_idx + first_idx] = q0 * cos_val - q1 * sin_val;
|
|
176
|
-
- input[q_base_idx + second_idx] = q0 * sin_val + q1 * cos_val;
|
|
177
|
-
-
|
|
178
|
-
- // Process K
|
|
179
|
-
- let k0 = input[k_base_idx + first_idx];
|
|
180
|
-
- let k1 = input[k_base_idx + second_idx];
|
|
181
|
-
- input[k_base_idx + first_idx] = k0 * cos_val - k1 * sin_val;
|
|
182
|
-
- input[k_base_idx + second_idx] = k0 * sin_val + k1 * cos_val;
|
|
183
|
-
+ let q0 = f32(input[q_base_idx + first_idx]);
|
|
184
|
-
+ let q1 = f32(input[q_base_idx + second_idx]);
|
|
185
|
-
+ input[q_base_idx + first_idx] = f16(q0 * cos_val - q1 * sin_val);
|
|
186
|
-
+ input[q_base_idx + second_idx] = f16(q0 * sin_val + q1 * cos_val);
|
|
187
|
-
+
|
|
188
|
-
+ let k0 = f32(input[k_base_idx + first_idx]);
|
|
189
|
-
+ let k1 = f32(input[k_base_idx + second_idx]);
|
|
190
|
-
+ input[k_base_idx + first_idx] = f16(k0 * cos_val - k1 * sin_val);
|
|
191
|
-
+ input[k_base_idx + second_idx] = f16(k0 * sin_val + k1 * cos_val);
|
|
192
|
-
}
|
|
193
|
-
|
|
194
|
-
// Precompute frequency table (run once at init)
|
|
195
|
-
-// Output: freqs_cos, freqs_sin [maxSeqLen, head_dim/2]
|
|
196
|
-
@compute @workgroup_size(WORKGROUP_SIZE, 1, 1)
|
|
197
|
-
fn precompute_freqs(
|
|
198
|
-
@builtin(global_invocation_id) global_id: vec3<u32>
|
|
199
|
-
) {
|
|
200
|
-
- let head_dim = u.head_dim;
|
|
201
|
-
- let seq_len = u.seq_len; // maxSeqLen for precomputation
|
|
202
|
-
+ let seq_len = u.seq_len;
|
|
203
|
-
let rope_base = u.rope_base;
|
|
204
|
-
let rope_scale = u.rope_scale;
|
|
205
|
-
let rotary_dim = u.rotary_dim;
|
|
206
|
-
@@ -245,8 +229,6 @@
|
|
207
|
-
freqs_sin[idx] = sin(theta);
|
|
208
|
-
}
|
|
209
|
-
|
|
210
|
-
-// NTK-aware scaled RoPE (for extended context without fine-tuning)
|
|
211
|
-
-// Uses dynamic scaling based on sequence length
|
|
212
|
-
@compute @workgroup_size(WORKGROUP_SIZE, 1, 1)
|
|
213
|
-
fn rope_ntk_scaled(
|
|
214
|
-
@builtin(global_invocation_id) global_id: vec3<u32>
|
|
215
|
-
@@ -268,8 +250,6 @@
|
|
216
|
-
return;
|
|
217
|
-
}
|
|
218
|
-
|
|
219
|
-
- // NTK scaling: increase base proportionally to scale factor
|
|
220
|
-
- // This preserves high-frequency components better than linear interpolation
|
|
221
|
-
rope_base = rope_base * pow(rope_scale, f32(pair_span_dim) / (f32(pair_span_dim) - 2.0));
|
|
222
|
-
|
|
223
|
-
let pos = idx / (num_heads * half_dim);
|
|
224
|
-
@@ -289,15 +269,13 @@
|
|
225
|
-
let base_idx = pos * num_heads * head_dim + head_idx * head_dim;
|
|
226
|
-
let first_idx = get_first_rotary_idx(pair_idx);
|
|
227
|
-
let second_idx = get_second_rotary_idx(pair_idx, pair_span_dim);
|
|
228
|
-
- let x0 = input[base_idx + first_idx];
|
|
229
|
-
- let x1 = input[base_idx + second_idx];
|
|
230
|
-
+ let x0 = f32(input[base_idx + first_idx]);
|
|
231
|
-
+ let x1 = f32(input[base_idx + second_idx]);
|
|
232
|
-
|
|
233
|
-
- input[base_idx + first_idx] = x0 * cos_val - x1 * sin_val;
|
|
234
|
-
- input[base_idx + second_idx] = x0 * sin_val + x1 * cos_val;
|
|
235
|
-
+ input[base_idx + first_idx] = f16(x0 * cos_val - x1 * sin_val);
|
|
236
|
-
+ input[base_idx + second_idx] = f16(x0 * sin_val + x1 * cos_val);
|
|
237
|
-
}
|
|
238
|
-
|
|
239
|
-
-// YaRN-style RoPE with attention scaling
|
|
240
|
-
-// Combines NTK interpolation with linear interpolation based on frequency
|
|
241
|
-
@compute @workgroup_size(WORKGROUP_SIZE, 1, 1)
|
|
242
|
-
fn rope_yarn(
|
|
243
|
-
@builtin(global_invocation_id) global_id: vec3<u32>
|
|
244
|
-
@@ -326,32 +304,26 @@
|
|
245
|
-
|
|
246
|
-
let actual_pos = f32(start_pos + pos);
|
|
247
|
-
|
|
248
|
-
- // YaRN parameters (from override constants)
|
|
249
|
-
let beta_fast: f32 = YARN_BETA_FAST;
|
|
250
|
-
let beta_slow: f32 = YARN_BETA_SLOW;
|
|
251
|
-
- let alpha: f32 = YARN_ALPHA;
|
|
252
|
-
|
|
253
|
-
- // Compute original frequency
|
|
254
|
-
let exponent = f32(pair_idx * 2u) / f32(pair_span_dim);
|
|
255
|
-
let orig_freq = 1.0 / pow(rope_base, exponent);
|
|
256
|
-
|
|
257
|
-
- // Compute wavelength
|
|
258
|
-
let wavelength = 2.0 * PI / orig_freq;
|
|
259
|
-
|
|
260
|
-
- // Interpolation factor based on wavelength
|
|
261
|
-
var ramp: f32;
|
|
262
|
-
let low_wavelength = f32(pair_span_dim) / beta_fast;
|
|
263
|
-
let high_wavelength = f32(pair_span_dim) / beta_slow;
|
|
264
|
-
|
|
265
|
-
if (wavelength < low_wavelength) {
|
|
266
|
-
- ramp = 0.0; // No interpolation for high frequencies
|
|
267
|
-
+ ramp = 0.0;
|
|
268
|
-
} else if (wavelength > high_wavelength) {
|
|
269
|
-
- ramp = 1.0; // Full interpolation for low frequencies
|
|
270
|
-
+ ramp = 1.0;
|
|
271
|
-
} else {
|
|
272
|
-
ramp = (wavelength - low_wavelength) / (high_wavelength - low_wavelength);
|
|
273
|
-
}
|
|
274
|
-
|
|
275
|
-
- // Combine original and scaled position
|
|
276
|
-
let scaled_pos = actual_pos / rope_scale;
|
|
277
|
-
let interp_pos = (1.0 - ramp) * actual_pos + ramp * scaled_pos;
|
|
278
|
-
|
|
279
|
-
@@ -362,9 +334,9 @@
|
|
280
|
-
let base_idx = pos * num_heads * head_dim + head_idx * head_dim;
|
|
281
|
-
let first_idx = get_first_rotary_idx(pair_idx);
|
|
282
|
-
let second_idx = get_second_rotary_idx(pair_idx, pair_span_dim);
|
|
283
|
-
- let x0 = input[base_idx + first_idx];
|
|
284
|
-
- let x1 = input[base_idx + second_idx];
|
|
285
|
-
+ let x0 = f32(input[base_idx + first_idx]);
|
|
286
|
-
+ let x1 = f32(input[base_idx + second_idx]);
|
|
287
|
-
|
|
288
|
-
- input[base_idx + first_idx] = x0 * cos_val - x1 * sin_val;
|
|
289
|
-
- input[base_idx + second_idx] = x0 * sin_val + x1 * cos_val;
|
|
290
|
-
+ input[base_idx + first_idx] = f16(x0 * cos_val - x1 * sin_val);
|
|
291
|
-
+ input[base_idx + second_idx] = f16(x0 * sin_val + x1 * cos_val);
|
|
292
|
-
}
|
|
@@ -1,336 +0,0 @@
|
|
|
1
|
-
--- src/gpu/kernels/sample.wgsl
|
|
2
|
-
+++ src/gpu/kernels/sample_f16.wgsl
|
|
3
|
-
@@ -1,47 +1,39 @@
|
|
4
|
-
-// sample.wgsl
|
|
5
|
-
+// sample_f16.wgsl
|
|
6
|
-
|
|
7
|
-
/**
|
|
8
|
-
- * GPU-Side Sampling Kernel
|
|
9
|
-
+ * GPU-Side Sampling Kernel (f16 logits)
|
|
10
|
-
*
|
|
11
|
-
- * Performs temperature scaling, top-k selection, softmax, and sampling
|
|
12
|
-
- * entirely on GPU. Only reads back the single selected token ID.
|
|
13
|
-
- *
|
|
14
|
-
- * Reduces readback from 1MB (256K vocab × 4 bytes) to 4 bytes.
|
|
15
|
-
- *
|
|
16
|
-
- * Algorithm:
|
|
17
|
-
- * 1. Temperature scaling: logits = logits / temperature
|
|
18
|
-
- * 2. Parallel top-k: Each workgroup finds local top-k, then merge
|
|
19
|
-
- * 3. Softmax on top-k candidates
|
|
20
|
-
- * 4. Multinomial sampling with provided random value
|
|
21
|
-
+ * Same as sample.wgsl but logits are f16.
|
|
22
|
-
+ * The all-f16 lane keeps sampling comparisons and reductions in f16.
|
|
23
|
-
*/
|
|
24
|
-
|
|
25
|
-
+enable f16;
|
|
26
|
-
+
|
|
27
|
-
// Configuration
|
|
28
|
-
override WORKGROUP_SIZE: u32 = 256u;
|
|
29
|
-
const MAX_WORKGROUP_SIZE: u32 = 256u;
|
|
30
|
-
-const MAX_TOP_K: u32 = 128u; // Max top-k supported
|
|
31
|
-
-const NEG_INF: f32 = -3.402823e+38;
|
|
32
|
-
+const MAX_TOP_K: u32 = 128u;
|
|
33
|
-
+const NEG_INF = f16(-65504.0);
|
|
34
|
-
|
|
35
|
-
struct Uniforms {
|
|
36
|
-
vocab_size: u32,
|
|
37
|
-
top_k: u32,
|
|
38
|
-
temperature: f32,
|
|
39
|
-
- random_value: f32, // Pre-generated random [0,1) for sampling
|
|
40
|
-
+ random_value: f32,
|
|
41
|
-
pad_token_id: u32,
|
|
42
|
-
- logit_softcap: f32, // Gemma 2: 30.0, 0.0 = disabled
|
|
43
|
-
- output_index: u32, // Index into output token buffer
|
|
44
|
-
+ logit_softcap: f32,
|
|
45
|
-
+ output_index: u32,
|
|
46
|
-
pad0: u32,
|
|
47
|
-
}
|
|
48
|
-
|
|
49
|
-
-// Apply softcapping: softcap * tanh(x / softcap)
|
|
50
|
-
-// Returns x unchanged if softcap <= 0
|
|
51
|
-
-fn apply_softcap(x: f32, softcap: f32) -> f32 {
|
|
52
|
-
- if (softcap <= 0.0) {
|
|
53
|
-
+fn apply_softcap(x: f16, softcap: f16) -> f16 {
|
|
54
|
-
+ if (softcap <= f16(0.0)) {
|
|
55
|
-
return x;
|
|
56
|
-
}
|
|
57
|
-
return softcap * tanh(x / softcap);
|
|
58
|
-
}
|
|
59
|
-
|
|
60
|
-
-fn candidate_beats(candidate_value: f32, candidate_index: u32, best_value: f32, best_index: u32) -> bool {
|
|
61
|
-
+fn candidate_beats(candidate_value: f16, candidate_index: u32, best_value: f16, best_index: u32) -> bool {
|
|
62
|
-
if (candidate_value > best_value) {
|
|
63
|
-
return true;
|
|
64
|
-
}
|
|
65
|
-
@@ -52,17 +44,14 @@
|
|
66
|
-
}
|
|
67
|
-
|
|
68
|
-
@group(0) @binding(0) var<uniform> u: Uniforms;
|
|
69
|
-
-@group(0) @binding(1) var<storage, read> logits: array<f32>; // [vocabSize]
|
|
70
|
-
-@group(0) @binding(2) var<storage, read_write> output: array<u32>; // [N] - selected tokens
|
|
71
|
-
-@group(0) @binding(3) var<storage, read_write> topk_indices: array<u32>; // [topK] - intermediate
|
|
72
|
-
-@group(0) @binding(4) var<storage, read_write> topk_logits: array<f32>; // [topK] - intermediate
|
|
73
|
-
+@group(0) @binding(1) var<storage, read> logits: array<f16>;
|
|
74
|
-
+@group(0) @binding(2) var<storage, read_write> output: array<u32>;
|
|
75
|
-
+@group(0) @binding(3) var<storage, read_write> topk_indices: array<u32>;
|
|
76
|
-
+@group(0) @binding(4) var<storage, read_write> topk_logits: array<f16>;
|
|
77
|
-
|
|
78
|
-
-// Shared memory for workgroup-level reduction
|
|
79
|
-
-var<workgroup> shared_values: array<f32, MAX_WORKGROUP_SIZE>;
|
|
80
|
-
+var<workgroup> shared_values: array<f16, MAX_WORKGROUP_SIZE>;
|
|
81
|
-
var<workgroup> shared_indices: array<u32, MAX_WORKGROUP_SIZE>;
|
|
82
|
-
|
|
83
|
-
-// Phase 1: Find local max in each workgroup for parallel top-k
|
|
84
|
-
-// Each thread scans a chunk of vocabulary, keeps local top element
|
|
85
|
-
@compute @workgroup_size(WORKGROUP_SIZE, 1, 1)
|
|
86
|
-
fn find_topk_phase1(
|
|
87
|
-
@builtin(global_invocation_id) gid: vec3<u32>,
|
|
88
|
-
@@ -73,13 +62,12 @@
|
|
89
|
-
let thread_idx = lid.x;
|
|
90
|
-
let global_idx = gid.x;
|
|
91
|
-
let vocab_size = u.vocab_size;
|
|
92
|
-
- let temperature = u.temperature;
|
|
93
|
-
+ let temperature = f16(u.temperature);
|
|
94
|
-
let pad_id = u.pad_token_id;
|
|
95
|
-
- let softcap = u.logit_softcap;
|
|
96
|
-
+ let softcap = f16(u.logit_softcap);
|
|
97
|
-
|
|
98
|
-
- // Single workgroup: write all logits directly for exact top-k
|
|
99
|
-
if (num_wg.x == 1u) {
|
|
100
|
-
- var val: f32 = NEG_INF;
|
|
101
|
-
+ var val: f16 = NEG_INF;
|
|
102
|
-
if (global_idx < vocab_size && global_idx != pad_id) {
|
|
103
|
-
val = apply_softcap(logits[global_idx], softcap) / temperature;
|
|
104
|
-
}
|
|
105
|
-
@@ -88,15 +76,12 @@
|
|
106
|
-
return;
|
|
107
|
-
}
|
|
108
|
-
|
|
109
|
-
- // Each thread finds max in its assigned range
|
|
110
|
-
- var local_max: f32 = NEG_INF; // -FLT_MAX
|
|
111
|
-
+ var local_max: f16 = NEG_INF;
|
|
112
|
-
var local_max_idx: u32 = 0u;
|
|
113
|
-
|
|
114
|
-
- // Stride through vocabulary
|
|
115
|
-
var idx = global_idx;
|
|
116
|
-
while (idx < vocab_size) {
|
|
117
|
-
if (idx != pad_id) {
|
|
118
|
-
- // Apply softcapping before temperature scaling
|
|
119
|
-
let val = apply_softcap(logits[idx], softcap) / temperature;
|
|
120
|
-
if (candidate_beats(val, idx, local_max, local_max_idx)) {
|
|
121
|
-
local_max = val;
|
|
122
|
-
@@ -110,7 +95,6 @@
|
|
123
|
-
shared_indices[thread_idx] = local_max_idx;
|
|
124
|
-
workgroupBarrier();
|
|
125
|
-
|
|
126
|
-
- // Reduce within workgroup to find workgroup's top value
|
|
127
|
-
var stride = WORKGROUP_SIZE / 2u;
|
|
128
|
-
while (stride > 0u) {
|
|
129
|
-
if (thread_idx < stride) {
|
|
130
|
-
@@ -128,15 +112,12 @@
|
|
131
|
-
stride = stride / 2u;
|
|
132
|
-
}
|
|
133
|
-
|
|
134
|
-
- // Thread 0 writes workgroup result
|
|
135
|
-
if (thread_idx == 0u) {
|
|
136
|
-
topk_logits[wgid.x] = shared_values[0];
|
|
137
|
-
topk_indices[wgid.x] = shared_indices[0];
|
|
138
|
-
}
|
|
139
|
-
}
|
|
140
|
-
|
|
141
|
-
-// Phase 2: Merge workgroup results and select final top-k
|
|
142
|
-
-// Single workgroup sorts and selects top-k from workgroup results
|
|
143
|
-
@compute @workgroup_size(WORKGROUP_SIZE, 1, 1)
|
|
144
|
-
fn find_topk_phase2(
|
|
145
|
-
@builtin(local_invocation_id) lid: vec3<u32>
|
|
146
|
-
@@ -146,8 +127,6 @@
|
|
147
|
-
let num_groups = min(WORKGROUP_SIZE, (u.vocab_size + WORKGROUP_SIZE - 1u) / WORKGROUP_SIZE);
|
|
148
|
-
let num_candidates = select(num_groups, min(u.vocab_size, WORKGROUP_SIZE), num_groups == 1u);
|
|
149
|
-
|
|
150
|
-
- // Load workgroup results into shared memory
|
|
151
|
-
- // Assume <= WORKGROUP_SIZE workgroups from phase 1
|
|
152
|
-
if (thread_idx < WORKGROUP_SIZE) {
|
|
153
|
-
if (thread_idx < num_candidates) {
|
|
154
|
-
shared_values[thread_idx] = topk_logits[thread_idx];
|
|
155
|
-
@@ -159,7 +138,6 @@
|
|
156
|
-
}
|
|
157
|
-
workgroupBarrier();
|
|
158
|
-
|
|
159
|
-
- // Thread 0 does partial selection sort for top-k
|
|
160
|
-
if (thread_idx == 0u) {
|
|
161
|
-
for (var k: u32 = 0u; k < top_k && k < num_candidates; k = k + 1u) {
|
|
162
|
-
var max_idx = k;
|
|
163
|
-
@@ -182,7 +160,6 @@
|
|
164
|
-
}
|
|
165
|
-
}
|
|
166
|
-
|
|
167
|
-
- // Write sorted top-k back
|
|
168
|
-
for (var k: u32 = 0u; k < top_k; k = k + 1u) {
|
|
169
|
-
topk_logits[k] = shared_values[k];
|
|
170
|
-
topk_indices[k] = shared_indices[k];
|
|
171
|
-
@@ -190,42 +167,36 @@
|
|
172
|
-
}
|
|
173
|
-
}
|
|
174
|
-
|
|
175
|
-
-// Phase 3: Softmax on top-k and sample
|
|
176
|
-
@compute @workgroup_size(WORKGROUP_SIZE, 1, 1)
|
|
177
|
-
fn softmax_and_sample(
|
|
178
|
-
@builtin(local_invocation_id) lid: vec3<u32>
|
|
179
|
-
) {
|
|
180
|
-
let thread_idx = lid.x;
|
|
181
|
-
let top_k = u.top_k;
|
|
182
|
-
- let random_val = u.random_value;
|
|
183
|
-
+ let random_val = f16(u.random_value);
|
|
184
|
-
|
|
185
|
-
- // Load top-k logits
|
|
186
|
-
if (thread_idx < top_k) {
|
|
187
|
-
shared_values[thread_idx] = topk_logits[thread_idx];
|
|
188
|
-
shared_indices[thread_idx] = topk_indices[thread_idx];
|
|
189
|
-
}
|
|
190
|
-
workgroupBarrier();
|
|
191
|
-
|
|
192
|
-
- // Thread 0 does softmax and sampling
|
|
193
|
-
if (thread_idx == 0u) {
|
|
194
|
-
- // Find max for numerical stability
|
|
195
|
-
- var max_val: f32 = shared_values[0];
|
|
196
|
-
+ var max_val: f16 = shared_values[0];
|
|
197
|
-
for (var i: u32 = 1u; i < top_k; i = i + 1u) {
|
|
198
|
-
max_val = max(max_val, shared_values[i]);
|
|
199
|
-
}
|
|
200
|
-
|
|
201
|
-
- // Compute exp and sum
|
|
202
|
-
- var exp_sum: f32 = 0.0;
|
|
203
|
-
+ var exp_sum: f16 = f16(0.0);
|
|
204
|
-
for (var i: u32 = 0u; i < top_k; i = i + 1u) {
|
|
205
|
-
let exp_val = exp(shared_values[i] - max_val);
|
|
206
|
-
shared_values[i] = exp_val;
|
|
207
|
-
exp_sum = exp_sum + exp_val;
|
|
208
|
-
}
|
|
209
|
-
|
|
210
|
-
- // Normalize to probabilities and sample
|
|
211
|
-
- let inv_sum = 1.0 / exp_sum;
|
|
212
|
-
- var cum_prob: f32 = 0.0;
|
|
213
|
-
- var selected_token: u32 = shared_indices[top_k - 1u]; // Default to last
|
|
214
|
-
+ let inv_sum = f16(1.0) / exp_sum;
|
|
215
|
-
+ var cum_prob: f16 = f16(0.0);
|
|
216
|
-
+ var selected_token: u32 = shared_indices[top_k - 1u];
|
|
217
|
-
|
|
218
|
-
for (var i: u32 = 0u; i < top_k; i = i + 1u) {
|
|
219
|
-
let prob = shared_values[i] * inv_sum;
|
|
220
|
-
@@ -240,8 +211,6 @@
|
|
221
|
-
}
|
|
222
|
-
}
|
|
223
|
-
|
|
224
|
-
-// Combined single-pass version for smaller vocabularies (<= 65536)
|
|
225
|
-
-// Uses hierarchical reduction within single kernel
|
|
226
|
-
@compute @workgroup_size(WORKGROUP_SIZE, 1, 1)
|
|
227
|
-
fn sample_single_pass(
|
|
228
|
-
@builtin(global_invocation_id) gid: vec3<u32>,
|
|
229
|
-
@@ -251,19 +220,16 @@
|
|
230
|
-
let thread_idx = lid.x;
|
|
231
|
-
let vocab_size = u.vocab_size;
|
|
232
|
-
let top_k = min(u.top_k, MAX_TOP_K);
|
|
233
|
-
- let temperature = u.temperature;
|
|
234
|
-
- let random_val = u.random_value;
|
|
235
|
-
+ let temperature = f16(u.temperature);
|
|
236
|
-
let pad_id = u.pad_token_id;
|
|
237
|
-
- let softcap = u.logit_softcap;
|
|
238
|
-
+ let softcap = f16(u.logit_softcap);
|
|
239
|
-
|
|
240
|
-
- // Phase 1: Find global max
|
|
241
|
-
- var local_max: f32 = NEG_INF;
|
|
242
|
-
+ var local_max: f16 = NEG_INF;
|
|
243
|
-
var local_max_idx: u32 = 0u;
|
|
244
|
-
|
|
245
|
-
var idx = gid.x;
|
|
246
|
-
while (idx < vocab_size) {
|
|
247
|
-
if (idx != pad_id) {
|
|
248
|
-
- // Apply softcapping before temperature scaling
|
|
249
|
-
let val = apply_softcap(logits[idx], softcap) / temperature;
|
|
250
|
-
if (candidate_beats(val, idx, local_max, local_max_idx)) {
|
|
251
|
-
local_max = val;
|
|
252
|
-
@@ -277,7 +243,6 @@
|
|
253
|
-
shared_indices[thread_idx] = local_max_idx;
|
|
254
|
-
workgroupBarrier();
|
|
255
|
-
|
|
256
|
-
- // Reduce to find workgroup max
|
|
257
|
-
var stride = WORKGROUP_SIZE / 2u;
|
|
258
|
-
while (stride > 0u) {
|
|
259
|
-
if (thread_idx < stride) {
|
|
260
|
-
@@ -295,18 +260,11 @@
|
|
261
|
-
stride = stride / 2u;
|
|
262
|
-
}
|
|
263
|
-
|
|
264
|
-
- // For single workgroup, thread 0 can do everything
|
|
265
|
-
if (thread_idx == 0u && num_wg.x == 1u) {
|
|
266
|
-
- // We have top-1, but need top-k
|
|
267
|
-
- // For small vocab, just do the full selection
|
|
268
|
-
- // This simplified version selects top-1 only (greedy)
|
|
269
|
-
- // Full top-k sampling requires multi-pass for large vocab
|
|
270
|
-
-
|
|
271
|
-
output[u.output_index] = shared_indices[0];
|
|
272
|
-
}
|
|
273
|
-
}
|
|
274
|
-
|
|
275
|
-
-// Greedy argmax for deterministic decoding (temperature=0 equivalent)
|
|
276
|
-
@compute @workgroup_size(WORKGROUP_SIZE, 1, 1)
|
|
277
|
-
fn argmax(
|
|
278
|
-
@builtin(global_invocation_id) gid: vec3<u32>,
|
|
279
|
-
@@ -318,16 +276,14 @@
|
|
280
|
-
let global_idx = gid.x;
|
|
281
|
-
let vocab_size = u.vocab_size;
|
|
282
|
-
let pad_id = u.pad_token_id;
|
|
283
|
-
- let softcap = u.logit_softcap;
|
|
284
|
-
+ let softcap = f16(u.logit_softcap);
|
|
285
|
-
|
|
286
|
-
- // Each thread finds max in its chunk
|
|
287
|
-
- var local_max: f32 = NEG_INF;
|
|
288
|
-
+ var local_max: f16 = NEG_INF;
|
|
289
|
-
var local_max_idx: u32 = 0u;
|
|
290
|
-
|
|
291
|
-
var idx = global_idx;
|
|
292
|
-
while (idx < vocab_size) {
|
|
293
|
-
if (idx != pad_id) {
|
|
294
|
-
- // Apply softcapping (argmax is greedy, no temperature)
|
|
295
|
-
let val = apply_softcap(logits[idx], softcap);
|
|
296
|
-
if (candidate_beats(val, idx, local_max, local_max_idx)) {
|
|
297
|
-
local_max = val;
|
|
298
|
-
@@ -341,7 +297,6 @@
|
|
299
|
-
shared_indices[thread_idx] = local_max_idx;
|
|
300
|
-
workgroupBarrier();
|
|
301
|
-
|
|
302
|
-
- // Reduce within workgroup
|
|
303
|
-
var stride = WORKGROUP_SIZE / 2u;
|
|
304
|
-
while (stride > 0u) {
|
|
305
|
-
if (thread_idx < stride) {
|
|
306
|
-
@@ -359,14 +314,12 @@
|
|
307
|
-
stride = stride / 2u;
|
|
308
|
-
}
|
|
309
|
-
|
|
310
|
-
- // Write workgroup result to global memory
|
|
311
|
-
if (thread_idx == 0u) {
|
|
312
|
-
topk_logits[wgid.x] = shared_values[0];
|
|
313
|
-
topk_indices[wgid.x] = shared_indices[0];
|
|
314
|
-
}
|
|
315
|
-
}
|
|
316
|
-
|
|
317
|
-
-// Final reduction for argmax across workgroups
|
|
318
|
-
@compute @workgroup_size(WORKGROUP_SIZE, 1, 1)
|
|
319
|
-
fn argmax_reduce(
|
|
320
|
-
@builtin(local_invocation_id) lid: vec3<u32>
|
|
321
|
-
@@ -374,7 +327,6 @@
|
|
322
|
-
let thread_idx = lid.x;
|
|
323
|
-
let num_groups = min(WORKGROUP_SIZE, (u.vocab_size + WORKGROUP_SIZE - 1u) / WORKGROUP_SIZE);
|
|
324
|
-
|
|
325
|
-
- // Load workgroup maxes (up to WORKGROUP_SIZE)
|
|
326
|
-
if (thread_idx < num_groups) {
|
|
327
|
-
shared_values[thread_idx] = topk_logits[thread_idx];
|
|
328
|
-
shared_indices[thread_idx] = topk_indices[thread_idx];
|
|
329
|
-
@@ -384,7 +336,6 @@
|
|
330
|
-
}
|
|
331
|
-
workgroupBarrier();
|
|
332
|
-
|
|
333
|
-
- // Reduce
|
|
334
|
-
var stride = WORKGROUP_SIZE / 2u;
|
|
335
|
-
while (stride > 0u) {
|
|
336
|
-
if (thread_idx < stride) {
|
package/src/gpu/kernels/codegen/patches/scatter_add_dynamic_f16.from.scatter_add_dynamic.diff
DELETED
|
@@ -1,65 +0,0 @@
|
|
|
1
|
-
--- src/gpu/kernels/scatter_add_dynamic.wgsl
|
|
2
|
-
+++ src/gpu/kernels/scatter_add_dynamic_f16.wgsl
|
|
3
|
-
@@ -1,31 +1,27 @@
|
|
4
|
-
-// scatter_add_dynamic.wgsl
|
|
5
|
-
+// scatter_add_dynamic_f16.wgsl
|
|
6
|
-
|
|
7
|
-
/**
|
|
8
|
-
- * Scatter-Add Kernel (dynamic layout)
|
|
9
|
-
- *
|
|
10
|
-
- * Alternative layout: expert outputs stored per-expert with token batching.
|
|
11
|
-
+ * Scatter-Add Kernel (dynamic layout, f16 outputs)
|
|
12
|
-
*/
|
|
13
|
-
+
|
|
14
|
-
+enable f16;
|
|
15
|
-
|
|
16
|
-
override WORKGROUP_SIZE: u32 = 256u;
|
|
17
|
-
|
|
18
|
-
-// Alternative layout: expert outputs stored per-expert with token batching
|
|
19
|
-
-// Layout: expertOutputs[expertIdx][batchedTokenIdx][hiddenSize]
|
|
20
|
-
-// This version handles dynamic token-to-expert mapping
|
|
21
|
-
struct DynamicUniforms {
|
|
22
|
-
- num_tokens: u32, // Total number of tokens
|
|
23
|
-
- hidden_size: u32, // Hidden dimension
|
|
24
|
-
- top_k: u32, // Number of experts per token
|
|
25
|
-
+ num_tokens: u32,
|
|
26
|
-
+ hidden_size: u32,
|
|
27
|
-
+ top_k: u32,
|
|
28
|
-
_pad: u32,
|
|
29
|
-
}
|
|
30
|
-
|
|
31
|
-
@group(0) @binding(0) var<uniform> u_dyn: DynamicUniforms;
|
|
32
|
-
-@group(0) @binding(1) var<storage, read> expert_outputs_flat: array<f32>; // Flattened expert outputs
|
|
33
|
-
-@group(0) @binding(2) var<storage, read> routing_indices: array<u32>; // [num_tokens, top_k] expert indices
|
|
34
|
-
-@group(0) @binding(3) var<storage, read> routing_weights: array<f32>; // [num_tokens, top_k] weights
|
|
35
|
-
-@group(0) @binding(4) var<storage, read> token_offsets: array<u32>; // Per-expert token offsets
|
|
36
|
-
-@group(0) @binding(5) var<storage, read_write> output_dyn: array<f32>; // [num_tokens, hidden_size]
|
|
37
|
-
+@group(0) @binding(1) var<storage, read> expert_outputs_flat: array<f16>;
|
|
38
|
-
+@group(0) @binding(2) var<storage, read> routing_indices: array<u32>;
|
|
39
|
-
+@group(0) @binding(3) var<storage, read> routing_weights: array<f32>;
|
|
40
|
-
+@group(0) @binding(4) var<storage, read> token_offsets: array<u32>;
|
|
41
|
-
+@group(0) @binding(5) var<storage, read_write> output_dyn: array<f16>;
|
|
42
|
-
|
|
43
|
-
-// Dynamic scatter with per-expert token offset lookup
|
|
44
|
-
@compute @workgroup_size(WORKGROUP_SIZE, 1, 1)
|
|
45
|
-
fn scatter_add_dynamic(@builtin(global_invocation_id) gid: vec3<u32>) {
|
|
46
|
-
let tid = gid.x;
|
|
47
|
-
@@ -44,16 +40,11 @@
|
|
48
|
-
let routing_base = token_idx * top_k;
|
|
49
|
-
|
|
50
|
-
for (var k: u32 = 0u; k < top_k; k = k + 1u) {
|
|
51
|
-
- let expert_idx = routing_indices[routing_base + k];
|
|
52
|
-
let weight = routing_weights[routing_base + k];
|
|
53
|
-
-
|
|
54
|
-
- // Look up where this token's data is stored for this expert
|
|
55
|
-
- // tokenOffsets[token_idx * top_k + k] gives the offset into expertOutputsFlat
|
|
56
|
-
let data_offset = token_offsets[routing_base + k];
|
|
57
|
-
let expert_data_idx = data_offset * hidden_size + dim_idx;
|
|
58
|
-
-
|
|
59
|
-
- sum = sum + weight * expert_outputs_flat[expert_data_idx];
|
|
60
|
-
+ sum = sum + weight * f32(expert_outputs_flat[expert_data_idx]);
|
|
61
|
-
}
|
|
62
|
-
|
|
63
|
-
- output_dyn[tid] = sum;
|
|
64
|
-
+ output_dyn[tid] = f16(sum);
|
|
65
|
-
}
|