doppler-gpu 0.4.7 → 0.4.9
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 +74 -0
- package/README.md +93 -144
- package/models/catalog.json +1589 -0
- package/package.json +44 -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-3-5-9b-f16-af32.json +474 -0
- package/src/config/conversion/qwen3/qwen-3-5-9b-q4k-ehaf16.json +509 -0
- 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/backward-registry.json +13 -2
- package/src/config/kernels/kernel-ref-digests.js +34 -21
- package/src/config/kernels/registry.json +1402 -2623
- 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-3-5-9b-metal-correctness.json +46 -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/schema/training-metrics.schema.js +4 -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/bridge/trainer-artifact-bridge.d.ts +161 -0
- package/src/experimental/bridge/trainer-artifact-bridge.js +573 -0
- package/src/experimental/orchestration/index.d.ts +10 -1
- package/src/experimental/training/attention-backward.js +123 -55
- package/src/experimental/training/autograd.js +101 -5
- package/src/experimental/training/clip.js +6 -1
- package/src/experimental/training/datasets/text-pairs.js +80 -8
- package/src/experimental/training/distillation/student-fixture.js +470 -114
- package/src/experimental/training/lora-pipeline.js +202 -27
- package/src/experimental/training/lora.js +10 -4
- package/src/experimental/training/optimizer.js +9 -1
- package/src/experimental/training/runner.js +18 -4
- package/src/experimental/training/suite.js +1 -1
- package/src/experimental/training/workloads.js +11 -0
- 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/adam.js +6 -4
- package/src/gpu/kernels/backward/causal_conv1d_silu_backward.d.ts +24 -0
- package/src/gpu/kernels/backward/causal_conv1d_silu_backward.js +74 -0
- package/src/gpu/kernels/backward/causal_conv1d_silu_backward.wgsl +59 -0
- package/src/gpu/kernels/backward/cross_entropy_backward.wgsl +4 -0
- package/src/gpu/kernels/backward/gated_delta_recurrent_backward.d.ts +39 -0
- package/src/gpu/kernels/backward/gated_delta_recurrent_backward.js +124 -0
- package/src/gpu/kernels/backward/gated_delta_recurrent_backward.wgsl +177 -0
- package/src/gpu/kernels/backward/gated_rmsnorm_backward.d.ts +32 -0
- package/src/gpu/kernels/backward/gated_rmsnorm_backward.js +90 -0
- package/src/gpu/kernels/backward/gated_rmsnorm_backward.wgsl +86 -0
- package/src/gpu/kernels/backward/index.d.ts +3 -0
- package/src/gpu/kernels/backward/index.js +3 -0
- package/src/gpu/kernels/backward/matmul_backward.wgsl +6 -6
- package/src/gpu/kernels/backward/matmul_backward_f16w.wgsl +82 -0
- package/src/gpu/kernels/backward/rmsnorm_backward.js +26 -4
- package/src/gpu/kernels/backward/rmsnorm_backward.wgsl +6 -3
- package/src/gpu/kernels/backward/utils.d.ts +12 -2
- package/src/gpu/kernels/backward/utils.js +9 -2
- package/src/gpu/kernels/cast_f16_to_f32.wgsl +6 -2
- 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 +26 -0
- package/src/gpu/kernels/index.js +18 -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 +10 -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.d.ts +1 -0
- package/src/gpu/kernels/silu.js +19 -4
- 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/embed.js +1 -1
- 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 +130 -10
- package/src/inference/pipelines/text/ffn/dense.js +241 -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 +326 -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/lora-apply.js +4 -4
- package/src/inference/pipelines/text/model-load.js +6 -0
- package/src/inference/pipelines/text/moe-cpu.js +1 -0
- package/src/inference/pipelines/text/moe-gpu.js +1 -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 +150 -55
- 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 +2 -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/node-converter.js +20 -5
- package/src/tooling/opfs-cache.d.ts +21 -1
- package/src/tooling/opfs-cache.js +168 -32
- package/src/tooling-experimental-exports.shared.d.ts +25 -0
- package/src/tooling-experimental-exports.shared.js +18 -0
- 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
|
@@ -1,213 +0,0 @@
|
|
|
1
|
-
--- src/gpu/kernels/attention_decode.wgsl
|
|
2
|
-
+++ src/gpu/kernels/attention_decode_f16.wgsl
|
|
3
|
-
@@ -1,28 +1,26 @@
|
|
4
|
-
-// Fused Multi-Head Attention Kernel
|
|
5
|
-
+// Fused Multi-Head Attention Kernel (decode, f16 QKV + f16 output)
|
|
6
|
-
//
|
|
7
|
-
-// Implements fused Q @ K^T → scale → mask → softmax → @ V
|
|
8
|
-
-// Uses tiled/blocked approach to avoid materializing full attention matrix.
|
|
9
|
-
-// Supports grouped query attention (GQA) where numKVHeads < numHeads.
|
|
10
|
-
-//
|
|
11
|
-
-// Based on Flash Attention principles adapted for WebGPU.
|
|
12
|
-
+// Same as attention_decode_f16kv.wgsl but Q/K/V are f16 and output is f16.
|
|
13
|
-
+
|
|
14
|
-
+enable f16;
|
|
15
|
-
|
|
16
|
-
const MAX_HEAD_DIM: u32 = 64u;
|
|
17
|
-
const MAX_WORKGROUP_SIZE: u32 = 256u;
|
|
18
|
-
|
|
19
|
-
-override DECODE_WORKGROUP_SIZE: u32 = 256u; // Decode kernel workgroup size (must match MAX_WORKGROUP_SIZE)
|
|
20
|
-
+override DECODE_WORKGROUP_SIZE: u32 = 256u;
|
|
21
|
-
|
|
22
|
-
struct Uniforms {
|
|
23
|
-
- num_heads: u32, // Number of query heads
|
|
24
|
-
- num_kv_heads: u32, // Number of KV heads (for GQA)
|
|
25
|
-
- head_dim: u32, // Dimension per head
|
|
26
|
-
- seq_len: u32, // Current sequence length (for KV)
|
|
27
|
-
- query_len: u32, // Query length (1 for decode, seq_len for prefill)
|
|
28
|
-
- scale: f32, // 1/sqrt(head_dim)
|
|
29
|
-
- is_causal: u32, // Apply causal mask (1 = yes)
|
|
30
|
-
- start_pos: u32, // Absolute position offset for causal masking
|
|
31
|
-
- attn_softcap: f32, // Gemma 2: 50.0, 0 = disabled
|
|
32
|
-
- sliding_window: u32, // Sliding window size (0 = disabled, >0 = window size)
|
|
33
|
-
- kv_len_source: u32, // 0 = use uniform seq_len, 1 = use buffer
|
|
34
|
-
+ num_heads: u32,
|
|
35
|
-
+ num_kv_heads: u32,
|
|
36
|
-
+ head_dim: u32,
|
|
37
|
-
+ seq_len: u32,
|
|
38
|
-
+ query_len: u32,
|
|
39
|
-
+ scale: f32,
|
|
40
|
-
+ is_causal: u32,
|
|
41
|
-
+ start_pos: u32,
|
|
42
|
-
+ attn_softcap: f32,
|
|
43
|
-
+ sliding_window: u32,
|
|
44
|
-
+ kv_len_source: u32,
|
|
45
|
-
kv_start: u32,
|
|
46
|
-
page_size: u32,
|
|
47
|
-
kv_layout: u32,
|
|
48
|
-
@@ -30,44 +28,28 @@
|
|
49
|
-
}
|
|
50
|
-
|
|
51
|
-
@group(0) @binding(0) var<uniform> u: Uniforms;
|
|
52
|
-
-@group(0) @binding(1) var<storage, read> Q: array<f32>; // [query_len, num_heads, head_dim]
|
|
53
|
-
-@group(0) @binding(2) var<storage, read> K: array<f32>; // [seq_len, num_kv_heads, head_dim]
|
|
54
|
-
-@group(0) @binding(3) var<storage, read> V: array<f32>; // [seq_len, num_kv_heads, head_dim]
|
|
55
|
-
-@group(0) @binding(4) var<storage, read_write> output: array<f32>; // [query_len, num_heads, head_dim]
|
|
56
|
-
+@group(0) @binding(1) var<storage, read> Q: array<f16>;
|
|
57
|
-
+@group(0) @binding(2) var<storage, read> K: array<f16>;
|
|
58
|
-
+@group(0) @binding(3) var<storage, read> V: array<f16>;
|
|
59
|
-
+@group(0) @binding(4) var<storage, read_write> output: array<f16>;
|
|
60
|
-
@group(0) @binding(5) var<storage, read> kv_len_buffer: array<u32>;
|
|
61
|
-
@group(0) @binding(6) var<storage, read> page_table: array<u32>;
|
|
62
|
-
|
|
63
|
-
-// Online softmax accumulators (per-thread)
|
|
64
|
-
-// Sized for 256 to support attention_decode workgroup size
|
|
65
|
-
var<workgroup> row_max: array<f32, MAX_WORKGROUP_SIZE>;
|
|
66
|
-
var<workgroup> row_sum: array<f32, MAX_WORKGROUP_SIZE>;
|
|
67
|
-
|
|
68
|
-
-// Get KV head index for grouped query attention
|
|
69
|
-
fn get_kv_head_idx(query_head_idx: u32) -> u32 {
|
|
70
|
-
- // GQA: multiple query heads share one KV head
|
|
71
|
-
let heads_per_kv = u.num_heads / u.num_kv_heads;
|
|
72
|
-
return query_head_idx / heads_per_kv;
|
|
73
|
-
}
|
|
74
|
-
|
|
75
|
-
-// Check if position should be masked (causal + sliding window attention)
|
|
76
|
-
fn is_masked(query_pos: u32, key_pos: u32) -> bool {
|
|
77
|
-
- // Compute absolute positions
|
|
78
|
-
let abs_query = query_pos + u.start_pos;
|
|
79
|
-
let abs_key = u.kv_start + key_pos;
|
|
80
|
-
-
|
|
81
|
-
- // Causal mask: query can only attend to keys at same or earlier positions
|
|
82
|
-
- if (u.is_causal != 0u && abs_key > abs_query) {
|
|
83
|
-
- return true;
|
|
84
|
-
- }
|
|
85
|
-
-
|
|
86
|
-
- // Sliding window mask: query can only attend to keys within the window
|
|
87
|
-
- // Key must be >= (query - window + 1), i.e., within the last `window` positions
|
|
88
|
-
+ if (u.is_causal != 0u && abs_key > abs_query) { return true; }
|
|
89
|
-
if (u.sliding_window > 0u && abs_query >= u.sliding_window) {
|
|
90
|
-
- if (abs_key < abs_query - u.sliding_window + 1u) {
|
|
91
|
-
- return true; // Key is too far in the past
|
|
92
|
-
- }
|
|
93
|
-
- }
|
|
94
|
-
-
|
|
95
|
-
+ if (abs_key < abs_query - u.sliding_window + 1u) { return true; }
|
|
96
|
-
+ }
|
|
97
|
-
return false;
|
|
98
|
-
}
|
|
99
|
-
|
|
100
|
-
@@ -92,12 +74,6 @@
|
|
101
|
-
return kv_len_buffer[0];
|
|
102
|
-
}
|
|
103
|
-
|
|
104
|
-
-// Main attention kernel - one workgroup per (query_block, head)
|
|
105
|
-
-// Workgroups are dispatched linearly as: numQueryBlocks * numHeads.
|
|
106
|
-
-// head_idx and query_block_idx are derived from workgroup_id.x.
|
|
107
|
-
-
|
|
108
|
-
-// Simplified single-query attention for decode step
|
|
109
|
-
-// More efficient when query_len == 1
|
|
110
|
-
@compute @workgroup_size(DECODE_WORKGROUP_SIZE, 1, 1)
|
|
111
|
-
fn attention_decode(
|
|
112
|
-
@builtin(global_invocation_id) global_id: vec3<u32>,
|
|
113
|
-
@@ -119,17 +95,14 @@
|
|
114
|
-
let seq_len = get_kv_len();
|
|
115
|
-
let scale = u.scale;
|
|
116
|
-
|
|
117
|
-
- // Each thread handles a subset of key positions
|
|
118
|
-
let keys_per_thread = (seq_len + DECODE_WORKGROUP_SIZE - 1u) / DECODE_WORKGROUP_SIZE;
|
|
119
|
-
|
|
120
|
-
- // Load query (single position)
|
|
121
|
-
var q_local: array<f32, MAX_HEAD_DIM>;
|
|
122
|
-
let q_offset = head_idx * head_dim;
|
|
123
|
-
for (var d: u32 = 0u; d < head_dim; d = d + 1u) {
|
|
124
|
-
- q_local[d] = Q[q_offset + d];
|
|
125
|
-
- }
|
|
126
|
-
-
|
|
127
|
-
- // Pass 1: local max
|
|
128
|
-
+ q_local[d] = f32(Q[q_offset + d]);
|
|
129
|
-
+ }
|
|
130
|
-
+
|
|
131
|
-
var local_max: f32 = -3.402823e+38;
|
|
132
|
-
let query_pos = 0u;
|
|
133
|
-
|
|
134
|
-
@@ -143,11 +116,10 @@
|
|
135
|
-
|
|
136
|
-
var score: f32 = 0.0;
|
|
137
|
-
for (var d: u32 = 0u; d < head_dim; d = d + 1u) {
|
|
138
|
-
- score = score + q_local[d] * K[k_offset + d];
|
|
139
|
-
+ score = score + q_local[d] * f32(K[k_offset + d]);
|
|
140
|
-
}
|
|
141
|
-
score = score * scale;
|
|
142
|
-
|
|
143
|
-
- // Gemma 2 attention softcapping
|
|
144
|
-
if (u.attn_softcap > 0.0) {
|
|
145
|
-
score = tanh(score / u.attn_softcap) * u.attn_softcap;
|
|
146
|
-
}
|
|
147
|
-
@@ -155,11 +127,9 @@
|
|
148
|
-
local_max = max(local_max, score);
|
|
149
|
-
}
|
|
150
|
-
|
|
151
|
-
- // Store local max for reduction
|
|
152
|
-
row_max[thread_idx] = local_max;
|
|
153
|
-
workgroupBarrier();
|
|
154
|
-
|
|
155
|
-
- // Parallel reduction to find global max
|
|
156
|
-
for (var stride: u32 = DECODE_WORKGROUP_SIZE / 2u; stride > 0u; stride = stride >> 1u) {
|
|
157
|
-
if (thread_idx < stride && thread_idx + stride < DECODE_WORKGROUP_SIZE) {
|
|
158
|
-
row_max[thread_idx] = max(row_max[thread_idx], row_max[thread_idx + stride]);
|
|
159
|
-
@@ -169,7 +139,6 @@
|
|
160
|
-
|
|
161
|
-
let global_max = row_max[0];
|
|
162
|
-
|
|
163
|
-
- // Pass 2: local sum and output accumulator
|
|
164
|
-
var local_sum: f32 = 0.0;
|
|
165
|
-
var local_out: array<f32, MAX_HEAD_DIM>;
|
|
166
|
-
for (var d: u32 = 0u; d < head_dim; d = d + 1u) {
|
|
167
|
-
@@ -187,7 +156,7 @@
|
|
168
|
-
|
|
169
|
-
var score: f32 = 0.0;
|
|
170
|
-
for (var d: u32 = 0u; d < head_dim; d = d + 1u) {
|
|
171
|
-
- score = score + q_local[d] * K[k_offset + d];
|
|
172
|
-
+ score = score + q_local[d] * f32(K[k_offset + d]);
|
|
173
|
-
}
|
|
174
|
-
score = score * scale;
|
|
175
|
-
|
|
176
|
-
@@ -198,14 +167,13 @@
|
|
177
|
-
let exp_score = exp(score - global_max);
|
|
178
|
-
local_sum = local_sum + exp_score;
|
|
179
|
-
for (var d: u32 = 0u; d < head_dim; d = d + 1u) {
|
|
180
|
-
- local_out[d] = local_out[d] + exp_score * V[v_offset + d];
|
|
181
|
-
+ local_out[d] = local_out[d] + exp_score * f32(V[v_offset + d]);
|
|
182
|
-
}
|
|
183
|
-
}
|
|
184
|
-
|
|
185
|
-
row_sum[thread_idx] = local_sum;
|
|
186
|
-
workgroupBarrier();
|
|
187
|
-
|
|
188
|
-
- // Parallel reduction for sum
|
|
189
|
-
for (var stride: u32 = DECODE_WORKGROUP_SIZE / 2u; stride > 0u; stride = stride >> 1u) {
|
|
190
|
-
if (thread_idx < stride && thread_idx + stride < DECODE_WORKGROUP_SIZE) {
|
|
191
|
-
row_sum[thread_idx] = row_sum[thread_idx] + row_sum[thread_idx + stride];
|
|
192
|
-
@@ -216,7 +184,6 @@
|
|
193
|
-
let global_sum = row_sum[0];
|
|
194
|
-
let inv_sum = select(0.0, 1.0 / global_sum, global_sum > 0.0);
|
|
195
|
-
|
|
196
|
-
- // Reduction for output (per-dimension)
|
|
197
|
-
let out_offset = head_idx * head_dim;
|
|
198
|
-
for (var d: u32 = 0u; d < head_dim; d = d + 1u) {
|
|
199
|
-
row_sum[thread_idx] = local_out[d];
|
|
200
|
-
@@ -226,8 +193,8 @@
|
|
201
|
-
for (var t: u32 = 0u; t < DECODE_WORKGROUP_SIZE; t = t + 1u) {
|
|
202
|
-
sum = sum + row_sum[t];
|
|
203
|
-
}
|
|
204
|
-
- output[out_offset + d] = sum * inv_sum;
|
|
205
|
-
- }
|
|
206
|
-
- workgroupBarrier();
|
|
207
|
-
- }
|
|
208
|
-
-}
|
|
209
|
-
+ output[out_offset + d] = f16(sum * inv_sum);
|
|
210
|
-
+ }
|
|
211
|
-
+ workgroupBarrier();
|
|
212
|
-
+ }
|
|
213
|
-
+}
|
|
@@ -1,89 +0,0 @@
|
|
|
1
|
-
--- src/gpu/kernels/attention_decode.wgsl
|
|
2
|
-
+++ src/gpu/kernels/attention_decode_f16kv.wgsl
|
|
3
|
-
@@ -1,10 +1,9 @@
|
|
4
|
-
-// Fused Multi-Head Attention Kernel
|
|
5
|
-
+// Fused Multi-Head Attention Kernel (f16 KV)
|
|
6
|
-
//
|
|
7
|
-
-// Implements fused Q @ K^T → scale → mask → softmax → @ V
|
|
8
|
-
-// Uses tiled/blocked approach to avoid materializing full attention matrix.
|
|
9
|
-
-// Supports grouped query attention (GQA) where numKVHeads < numHeads.
|
|
10
|
-
-//
|
|
11
|
-
-// Based on Flash Attention principles adapted for WebGPU.
|
|
12
|
-
+// Same as attention.wgsl but K/V are stored as f16 for KV-cache compression.
|
|
13
|
-
+// Q and output remain f32; K/V are cast to f32 on load.
|
|
14
|
-
+
|
|
15
|
-
+enable f16;
|
|
16
|
-
|
|
17
|
-
const MAX_HEAD_DIM: u32 = 64u;
|
|
18
|
-
const MAX_WORKGROUP_SIZE: u32 = 256u;
|
|
19
|
-
@@ -31,8 +30,8 @@
|
|
20
|
-
|
|
21
|
-
@group(0) @binding(0) var<uniform> u: Uniforms;
|
|
22
|
-
@group(0) @binding(1) var<storage, read> Q: array<f32>; // [query_len, num_heads, head_dim]
|
|
23
|
-
-@group(0) @binding(2) var<storage, read> K: array<f32>; // [seq_len, num_kv_heads, head_dim]
|
|
24
|
-
-@group(0) @binding(3) var<storage, read> V: array<f32>; // [seq_len, num_kv_heads, head_dim]
|
|
25
|
-
+@group(0) @binding(2) var<storage, read> K: array<f16>; // [seq_len, num_kv_heads, head_dim]
|
|
26
|
-
+@group(0) @binding(3) var<storage, read> V: array<f16>; // [seq_len, num_kv_heads, head_dim]
|
|
27
|
-
@group(0) @binding(4) var<storage, read_write> output: array<f32>; // [query_len, num_heads, head_dim]
|
|
28
|
-
@group(0) @binding(5) var<storage, read> kv_len_buffer: array<u32>;
|
|
29
|
-
@group(0) @binding(6) var<storage, read> page_table: array<u32>;
|
|
30
|
-
@@ -51,23 +50,14 @@
|
|
31
|
-
|
|
32
|
-
// Check if position should be masked (causal + sliding window attention)
|
|
33
|
-
fn is_masked(query_pos: u32, key_pos: u32) -> bool {
|
|
34
|
-
- // Compute absolute positions
|
|
35
|
-
let abs_query = query_pos + u.start_pos;
|
|
36
|
-
let abs_key = u.kv_start + key_pos;
|
|
37
|
-
-
|
|
38
|
-
- // Causal mask: query can only attend to keys at same or earlier positions
|
|
39
|
-
- if (u.is_causal != 0u && abs_key > abs_query) {
|
|
40
|
-
- return true;
|
|
41
|
-
- }
|
|
42
|
-
-
|
|
43
|
-
- // Sliding window mask: query can only attend to keys within the window
|
|
44
|
-
- // Key must be >= (query - window + 1), i.e., within the last `window` positions
|
|
45
|
-
+ // Causal mask
|
|
46
|
-
+ if (u.is_causal != 0u && abs_key > abs_query) { return true; }
|
|
47
|
-
+ // Sliding window mask
|
|
48
|
-
if (u.sliding_window > 0u && abs_query >= u.sliding_window) {
|
|
49
|
-
- if (abs_key < abs_query - u.sliding_window + 1u) {
|
|
50
|
-
- return true; // Key is too far in the past
|
|
51
|
-
- }
|
|
52
|
-
- }
|
|
53
|
-
-
|
|
54
|
-
+ if (abs_key < abs_query - u.sliding_window + 1u) { return true; }
|
|
55
|
-
+ }
|
|
56
|
-
return false;
|
|
57
|
-
}
|
|
58
|
-
|
|
59
|
-
@@ -143,11 +133,10 @@
|
|
60
|
-
|
|
61
|
-
var score: f32 = 0.0;
|
|
62
|
-
for (var d: u32 = 0u; d < head_dim; d = d + 1u) {
|
|
63
|
-
- score = score + q_local[d] * K[k_offset + d];
|
|
64
|
-
+ score = score + q_local[d] * f32(K[k_offset + d]);
|
|
65
|
-
}
|
|
66
|
-
score = score * scale;
|
|
67
|
-
|
|
68
|
-
- // Gemma 2 attention softcapping
|
|
69
|
-
if (u.attn_softcap > 0.0) {
|
|
70
|
-
score = tanh(score / u.attn_softcap) * u.attn_softcap;
|
|
71
|
-
}
|
|
72
|
-
@@ -187,7 +176,7 @@
|
|
73
|
-
|
|
74
|
-
var score: f32 = 0.0;
|
|
75
|
-
for (var d: u32 = 0u; d < head_dim; d = d + 1u) {
|
|
76
|
-
- score = score + q_local[d] * K[k_offset + d];
|
|
77
|
-
+ score = score + q_local[d] * f32(K[k_offset + d]);
|
|
78
|
-
}
|
|
79
|
-
score = score * scale;
|
|
80
|
-
|
|
81
|
-
@@ -198,7 +187,7 @@
|
|
82
|
-
let exp_score = exp(score - global_max);
|
|
83
|
-
local_sum = local_sum + exp_score;
|
|
84
|
-
for (var d: u32 = 0u; d < head_dim; d = d + 1u) {
|
|
85
|
-
- local_out[d] = local_out[d] + exp_score * V[v_offset + d];
|
|
86
|
-
+ local_out[d] = local_out[d] + exp_score * f32(V[v_offset + d]);
|
|
87
|
-
}
|
|
88
|
-
}
|
|
89
|
-
|
|
@@ -1,62 +0,0 @@
|
|
|
1
|
-
--- src/gpu/kernels/attention.wgsl
|
|
2
|
-
+++ src/gpu/kernels/attention_f16.wgsl
|
|
3
|
-
@@ -6,6 +6,8 @@
|
|
4
|
-
//
|
|
5
|
-
// Based on Flash Attention principles adapted for WebGPU.
|
|
6
|
-
|
|
7
|
-
+enable f16;
|
|
8
|
-
+
|
|
9
|
-
// Tile sizes for blocked attention
|
|
10
|
-
const MAX_BLOCK_SIZE: u32 = 32u;
|
|
11
|
-
const MAX_HEAD_TILE: u32 = 64u;
|
|
12
|
-
@@ -38,10 +40,10 @@
|
|
13
|
-
}
|
|
14
|
-
|
|
15
|
-
@group(0) @binding(0) var<uniform> u: Uniforms;
|
|
16
|
-
-@group(0) @binding(1) var<storage, read> Q: array<f32>; // [query_len, num_heads, head_dim]
|
|
17
|
-
-@group(0) @binding(2) var<storage, read> K: array<f32>; // [seq_len, num_kv_heads, head_dim]
|
|
18
|
-
-@group(0) @binding(3) var<storage, read> V: array<f32>; // [seq_len, num_kv_heads, head_dim]
|
|
19
|
-
-@group(0) @binding(4) var<storage, read_write> output: array<f32>; // [query_len, num_heads, head_dim]
|
|
20
|
-
+@group(0) @binding(1) var<storage, read> Q: array<f16>; // [query_len, num_heads, head_dim]
|
|
21
|
-
+@group(0) @binding(2) var<storage, read> K: array<f16>; // [seq_len, num_kv_heads, head_dim]
|
|
22
|
-
+@group(0) @binding(3) var<storage, read> V: array<f16>; // [seq_len, num_kv_heads, head_dim]
|
|
23
|
-
+@group(0) @binding(4) var<storage, read_write> output: array<f16>; // [query_len, num_heads, head_dim]
|
|
24
|
-
@group(0) @binding(5) var<storage, read> kv_len_buffer: array<u32>;
|
|
25
|
-
@group(0) @binding(6) var<storage, read> page_table: array<u32>;
|
|
26
|
-
|
|
27
|
-
@@ -156,7 +158,7 @@
|
|
28
|
-
if (valid_query) {
|
|
29
|
-
let q_offset = query_pos * u.num_heads * head_dim + head_idx * head_dim;
|
|
30
|
-
for (var d: u32 = 0u; d < head_dim; d = d + 1u) {
|
|
31
|
-
- q_local[d] = Q[q_offset + d];
|
|
32
|
-
+ q_local[d] = f32(Q[q_offset + d]);
|
|
33
|
-
}
|
|
34
|
-
}
|
|
35
|
-
|
|
36
|
-
@@ -172,7 +174,7 @@
|
|
37
|
-
let k_idx = get_kv_pos(k_load_idx);
|
|
38
|
-
let k_offset = k_idx * u.num_kv_heads * head_dim + kv_head_idx * head_dim;
|
|
39
|
-
for (var d: u32 = 0u; d < head_dim; d = d + 1u) {
|
|
40
|
-
- shared_K[thread_idx * head_dim + d] = K[k_offset + d];
|
|
41
|
-
+ shared_K[thread_idx * head_dim + d] = f32(K[k_offset + d]);
|
|
42
|
-
}
|
|
43
|
-
} else {
|
|
44
|
-
for (var d: u32 = 0u; d < head_dim; d = d + 1u) {
|
|
45
|
-
@@ -186,7 +188,7 @@
|
|
46
|
-
let v_idx = get_kv_pos(v_load_idx);
|
|
47
|
-
let v_offset = v_idx * u.num_kv_heads * head_dim + kv_head_idx * head_dim;
|
|
48
|
-
for (var d: u32 = 0u; d < head_dim; d = d + 1u) {
|
|
49
|
-
- shared_V[thread_idx * head_dim + d] = V[v_offset + d];
|
|
50
|
-
+ shared_V[thread_idx * head_dim + d] = f32(V[v_offset + d]);
|
|
51
|
-
}
|
|
52
|
-
} else {
|
|
53
|
-
for (var d: u32 = 0u; d < head_dim; d = d + 1u) {
|
|
54
|
-
@@ -262,7 +264,7 @@
|
|
55
|
-
let out_offset = query_pos * u.num_heads * head_dim + head_idx * head_dim;
|
|
56
|
-
let inv_l_i = select(0.0, 1.0 / l_i, l_i > 0.0);
|
|
57
|
-
for (var d: u32 = 0u; d < head_dim; d = d + 1u) {
|
|
58
|
-
- output[out_offset + d] = acc[d] * inv_l_i;
|
|
59
|
-
+ output[out_offset + d] = f16(acc[d] * inv_l_i);
|
|
60
|
-
}
|
|
61
|
-
}
|
|
62
|
-
}
|
|
@@ -1,40 +0,0 @@
|
|
|
1
|
-
--- src/gpu/kernels/attention.wgsl
|
|
2
|
-
+++ src/gpu/kernels/attention_f16kv.wgsl
|
|
3
|
-
@@ -6,6 +6,8 @@
|
|
4
|
-
//
|
|
5
|
-
// Based on Flash Attention principles adapted for WebGPU.
|
|
6
|
-
|
|
7
|
-
+enable f16;
|
|
8
|
-
+
|
|
9
|
-
// Tile sizes for blocked attention
|
|
10
|
-
const MAX_BLOCK_SIZE: u32 = 32u;
|
|
11
|
-
const MAX_HEAD_TILE: u32 = 64u;
|
|
12
|
-
@@ -39,8 +41,8 @@
|
|
13
|
-
|
|
14
|
-
@group(0) @binding(0) var<uniform> u: Uniforms;
|
|
15
|
-
@group(0) @binding(1) var<storage, read> Q: array<f32>; // [query_len, num_heads, head_dim]
|
|
16
|
-
-@group(0) @binding(2) var<storage, read> K: array<f32>; // [seq_len, num_kv_heads, head_dim]
|
|
17
|
-
-@group(0) @binding(3) var<storage, read> V: array<f32>; // [seq_len, num_kv_heads, head_dim]
|
|
18
|
-
+@group(0) @binding(2) var<storage, read> K: array<f16>; // [seq_len, num_kv_heads, head_dim]
|
|
19
|
-
+@group(0) @binding(3) var<storage, read> V: array<f16>; // [seq_len, num_kv_heads, head_dim]
|
|
20
|
-
@group(0) @binding(4) var<storage, read_write> output: array<f32>; // [query_len, num_heads, head_dim]
|
|
21
|
-
@group(0) @binding(5) var<storage, read> kv_len_buffer: array<u32>;
|
|
22
|
-
@group(0) @binding(6) var<storage, read> page_table: array<u32>;
|
|
23
|
-
@@ -172,7 +174,7 @@
|
|
24
|
-
let k_idx = get_kv_pos(k_load_idx);
|
|
25
|
-
let k_offset = k_idx * u.num_kv_heads * head_dim + kv_head_idx * head_dim;
|
|
26
|
-
for (var d: u32 = 0u; d < head_dim; d = d + 1u) {
|
|
27
|
-
- shared_K[thread_idx * head_dim + d] = K[k_offset + d];
|
|
28
|
-
+ shared_K[thread_idx * head_dim + d] = f32(K[k_offset + d]);
|
|
29
|
-
}
|
|
30
|
-
} else {
|
|
31
|
-
for (var d: u32 = 0u; d < head_dim; d = d + 1u) {
|
|
32
|
-
@@ -186,7 +188,7 @@
|
|
33
|
-
let v_idx = get_kv_pos(v_load_idx);
|
|
34
|
-
let v_offset = v_idx * u.num_kv_heads * head_dim + kv_head_idx * head_dim;
|
|
35
|
-
for (var d: u32 = 0u; d < head_dim; d = d + 1u) {
|
|
36
|
-
- shared_V[thread_idx * head_dim + d] = V[v_offset + d];
|
|
37
|
-
+ shared_V[thread_idx * head_dim + d] = f32(V[v_offset + d]);
|
|
38
|
-
}
|
|
39
|
-
} else {
|
|
40
|
-
for (var d: u32 = 0u; d < head_dim; d = d + 1u) {
|
|
@@ -1,226 +0,0 @@
|
|
|
1
|
-
--- src/gpu/kernels/attention_small.wgsl
|
|
2
|
-
+++ src/gpu/kernels/attention_small_f16.wgsl
|
|
3
|
-
@@ -1,39 +1,28 @@
|
|
4
|
-
-// Fused Multi-Head Attention Kernel (small tiles)
|
|
5
|
-
+// Fused Multi-Head Attention Kernel (small tiles, f16 QKV + f16 output)
|
|
6
|
-
//
|
|
7
|
-
-// Compatibility-focused variant for devices with limited workgroup storage
|
|
8
|
-
-// and for models with larger head_dim (up to MAX_HEAD_DIM).
|
|
9
|
-
-//
|
|
10
|
-
-// Differences vs attention.wgsl:
|
|
11
|
-
-// - BLOCK_SIZE = 32, HEAD_TILE = 32
|
|
12
|
-
-// - No shared score matrix (recompute in second pass)
|
|
13
|
-
-// - Tiles over head_dim to support large heads
|
|
14
|
-
-//
|
|
15
|
-
-// Q and output are f32. K/V are f32.
|
|
16
|
-
-//
|
|
17
|
-
-// One workgroup per (query_block, head) encoded linearly in workgroup_id.x:
|
|
18
|
-
-// linear = workgroup_id.x
|
|
19
|
-
-// head_idx = linear % num_heads
|
|
20
|
-
-// query_block_idx = linear / num_heads
|
|
21
|
-
-//
|
|
22
|
-
-// workgroup_size = BLOCK_SIZE (one thread per query position in the block).
|
|
23
|
-
+// Same algorithm as attention_small.wgsl but Q/K/V are f16 and output is f16.
|
|
24
|
-
+// Shared tiles and attention arithmetic use f16 for the all-f16 lane.
|
|
25
|
-
|
|
26
|
-
-const BLOCK_SIZE: u32 = 32u; // Sequence tile size
|
|
27
|
-
-const HEAD_TILE: u32 = 32u; // Head dimension tile
|
|
28
|
-
-const MAX_HEAD_DIM: u32 = 256u;
|
|
29
|
-
+enable f16;
|
|
30
|
-
+
|
|
31
|
-
+const BLOCK_SIZE: u32 = 32u;
|
|
32
|
-
const WORKGROUP_SIZE: u32 = BLOCK_SIZE;
|
|
33
|
-
+const HEAD_TILE: u32 = 32u;
|
|
34
|
-
+const MAX_HEAD_DIM: u32 = 256u;
|
|
35
|
-
+const NEG_INF = f16(-65504.0);
|
|
36
|
-
|
|
37
|
-
struct Uniforms {
|
|
38
|
-
- num_heads: u32, // Number of query heads
|
|
39
|
-
- num_kv_heads: u32, // Number of KV heads (for GQA)
|
|
40
|
-
- head_dim: u32, // Dimension per head
|
|
41
|
-
- seq_len: u32, // KV length
|
|
42
|
-
- query_len: u32, // Query length (1 for decode, >1 for prefill)
|
|
43
|
-
- scale: f32, // 1/sqrt(head_dim)
|
|
44
|
-
- is_causal: u32, // Apply causal mask (1 = yes)
|
|
45
|
-
- start_pos: u32, // Absolute position offset for causal masking
|
|
46
|
-
- attn_softcap: f32, // Gemma 2: 50.0, 0 = disabled
|
|
47
|
-
- sliding_window: u32, // Sliding window size (0 = disabled, >0 = window size)
|
|
48
|
-
- kv_len_source: u32, // 0 = use uniform seq_len, 1 = use buffer
|
|
49
|
-
+ num_heads: u32,
|
|
50
|
-
+ num_kv_heads: u32,
|
|
51
|
-
+ head_dim: u32,
|
|
52
|
-
+ seq_len: u32,
|
|
53
|
-
+ query_len: u32,
|
|
54
|
-
+ scale: f32,
|
|
55
|
-
+ is_causal: u32,
|
|
56
|
-
+ start_pos: u32,
|
|
57
|
-
+ attn_softcap: f32,
|
|
58
|
-
+ sliding_window: u32,
|
|
59
|
-
+ kv_len_source: u32,
|
|
60
|
-
kv_start: u32,
|
|
61
|
-
page_size: u32,
|
|
62
|
-
kv_layout: u32,
|
|
63
|
-
@@ -41,16 +30,15 @@
|
|
64
|
-
}
|
|
65
|
-
|
|
66
|
-
@group(0) @binding(0) var<uniform> u: Uniforms;
|
|
67
|
-
-@group(0) @binding(1) var<storage, read> Q: array<f32>; // [query_len, num_heads, head_dim]
|
|
68
|
-
-@group(0) @binding(2) var<storage, read> K: array<f32>; // [seq_len, num_kv_heads, head_dim]
|
|
69
|
-
-@group(0) @binding(3) var<storage, read> V: array<f32>; // [seq_len, num_kv_heads, head_dim]
|
|
70
|
-
-@group(0) @binding(4) var<storage, read_write> output: array<f32>; // [query_len, num_heads, head_dim]
|
|
71
|
-
+@group(0) @binding(1) var<storage, read> Q: array<f16>;
|
|
72
|
-
+@group(0) @binding(2) var<storage, read> K: array<f16>;
|
|
73
|
-
+@group(0) @binding(3) var<storage, read> V: array<f16>;
|
|
74
|
-
+@group(0) @binding(4) var<storage, read_write> output: array<f16>;
|
|
75
|
-
@group(0) @binding(5) var<storage, read> kv_len_buffer: array<u32>;
|
|
76
|
-
@group(0) @binding(6) var<storage, read> page_table: array<u32>;
|
|
77
|
-
|
|
78
|
-
-// Shared memory for tiled K/V slices
|
|
79
|
-
-var<workgroup> shared_K: array<f32, BLOCK_SIZE * HEAD_TILE>;
|
|
80
|
-
-var<workgroup> shared_V: array<f32, BLOCK_SIZE * HEAD_TILE>;
|
|
81
|
-
+var<workgroup> shared_K: array<f16, BLOCK_SIZE * HEAD_TILE>;
|
|
82
|
-
+var<workgroup> shared_V: array<f16, BLOCK_SIZE * HEAD_TILE>;
|
|
83
|
-
|
|
84
|
-
fn get_kv_head_idx(query_head_idx: u32) -> u32 {
|
|
85
|
-
let heads_per_kv = u.num_heads / u.num_kv_heads;
|
|
86
|
-
@@ -60,9 +48,7 @@
|
|
87
|
-
fn is_masked(query_pos: u32, key_pos: u32) -> bool {
|
|
88
|
-
let abs_query = query_pos + u.start_pos;
|
|
89
|
-
let abs_key = u.kv_start + key_pos;
|
|
90
|
-
- // Causal mask
|
|
91
|
-
if (u.is_causal != 0u && abs_key > abs_query) { return true; }
|
|
92
|
-
- // Sliding window mask
|
|
93
|
-
if (u.sliding_window > 0u && abs_query >= u.sliding_window) {
|
|
94
|
-
if (abs_key < abs_query - u.sliding_window + 1u) { return true; }
|
|
95
|
-
}
|
|
96
|
-
@@ -108,24 +94,25 @@
|
|
97
|
-
}
|
|
98
|
-
let seq_len = get_kv_len();
|
|
99
|
-
let query_len = u.query_len;
|
|
100
|
-
- let scale = u.scale;
|
|
101
|
-
+ let scale = f16(u.scale);
|
|
102
|
-
+ let softcap = f16(u.attn_softcap);
|
|
103
|
-
|
|
104
|
-
let query_pos = query_block_idx * BLOCK_SIZE + thread_idx;
|
|
105
|
-
let valid_query = query_pos < query_len;
|
|
106
|
-
|
|
107
|
-
- var q_local: array<f32, MAX_HEAD_DIM>;
|
|
108
|
-
- var acc: array<f32, MAX_HEAD_DIM>;
|
|
109
|
-
+ var q_local: array<f16, MAX_HEAD_DIM>;
|
|
110
|
-
+ var acc: array<f16, MAX_HEAD_DIM>;
|
|
111
|
-
|
|
112
|
-
if (valid_query) {
|
|
113
|
-
let q_offset = query_pos * num_heads * head_dim + head_idx * head_dim;
|
|
114
|
-
for (var d: u32 = 0u; d < head_dim; d = d + 1u) {
|
|
115
|
-
q_local[d] = Q[q_offset + d];
|
|
116
|
-
- acc[d] = 0.0;
|
|
117
|
-
+ acc[d] = f16(0.0);
|
|
118
|
-
}
|
|
119
|
-
}
|
|
120
|
-
|
|
121
|
-
- var m_i: f32 = -3.402823e+38;
|
|
122
|
-
- var l_i: f32 = 0.0;
|
|
123
|
-
+ var m_i: f16 = NEG_INF;
|
|
124
|
-
+ var l_i: f16 = f16(0.0);
|
|
125
|
-
|
|
126
|
-
let num_kv_blocks = (seq_len + BLOCK_SIZE - 1u) / BLOCK_SIZE;
|
|
127
|
-
let num_head_tiles = (head_dim + HEAD_TILE - 1u) / HEAD_TILE;
|
|
128
|
-
@@ -133,17 +120,15 @@
|
|
129
|
-
for (var kv_block: u32 = 0u; kv_block < num_kv_blocks; kv_block = kv_block + 1u) {
|
|
130
|
-
let kv_block_start = kv_block * BLOCK_SIZE;
|
|
131
|
-
|
|
132
|
-
- var scores: array<f32, BLOCK_SIZE>;
|
|
133
|
-
+ var scores: array<f16, BLOCK_SIZE>;
|
|
134
|
-
for (var k_init: u32 = 0u; k_init < BLOCK_SIZE; k_init = k_init + 1u) {
|
|
135
|
-
- scores[k_init] = 0.0;
|
|
136
|
-
+ scores[k_init] = f16(0.0);
|
|
137
|
-
}
|
|
138
|
-
|
|
139
|
-
- // Compute dot products for this KV block by tiling head_dim.
|
|
140
|
-
for (var ht: u32 = 0u; ht < num_head_tiles; ht = ht + 1u) {
|
|
141
|
-
let d0 = ht * HEAD_TILE;
|
|
142
|
-
let tile_len = min(HEAD_TILE, head_dim - d0);
|
|
143
|
-
|
|
144
|
-
- // Load K slice for this block into shared memory.
|
|
145
|
-
let key_pos_load = kv_block_start + thread_idx;
|
|
146
|
-
if (key_pos_load < seq_len) {
|
|
147
|
-
let k_idx = get_kv_pos(key_pos_load);
|
|
148
|
-
@@ -153,7 +138,7 @@
|
|
149
|
-
}
|
|
150
|
-
} else {
|
|
151
|
-
for (var td: u32 = 0u; td < tile_len; td = td + 1u) {
|
|
152
|
-
- shared_K[thread_idx * HEAD_TILE + td] = 0.0;
|
|
153
|
-
+ shared_K[thread_idx * HEAD_TILE + td] = f16(0.0);
|
|
154
|
-
}
|
|
155
|
-
}
|
|
156
|
-
|
|
157
|
-
@@ -165,7 +150,7 @@
|
|
158
|
-
if (key_pos >= seq_len) { continue; }
|
|
159
|
-
if (is_masked(query_pos, key_pos)) { continue; }
|
|
160
|
-
|
|
161
|
-
- var dot_partial: f32 = 0.0;
|
|
162
|
-
+ var dot_partial: f16 = f16(0.0);
|
|
163
|
-
for (var td: u32 = 0u; td < tile_len; td = td + 1u) {
|
|
164
|
-
dot_partial = dot_partial + q_local[d0 + td] * shared_K[k * HEAD_TILE + td];
|
|
165
|
-
}
|
|
166
|
-
@@ -176,18 +161,17 @@
|
|
167
|
-
workgroupBarrier();
|
|
168
|
-
}
|
|
169
|
-
|
|
170
|
-
- var m_new: f32 = m_i;
|
|
171
|
-
+ var m_new: f16 = m_i;
|
|
172
|
-
if (valid_query) {
|
|
173
|
-
- var block_max: f32 = -3.402823e+38;
|
|
174
|
-
+ var block_max: f16 = NEG_INF;
|
|
175
|
-
for (var k: u32 = 0u; k < BLOCK_SIZE; k = k + 1u) {
|
|
176
|
-
let key_pos = kv_block_start + k;
|
|
177
|
-
if (key_pos >= seq_len) { continue; }
|
|
178
|
-
if (is_masked(query_pos, key_pos)) { continue; }
|
|
179
|
-
|
|
180
|
-
var s = scores[k] * scale;
|
|
181
|
-
- // Gemma 2 attention softcapping
|
|
182
|
-
- if (u.attn_softcap > 0.0) {
|
|
183
|
-
- s = tanh(s / u.attn_softcap) * u.attn_softcap;
|
|
184
|
-
+ if (softcap > f16(0.0)) {
|
|
185
|
-
+ s = tanh(s / softcap) * softcap;
|
|
186
|
-
}
|
|
187
|
-
scores[k] = s;
|
|
188
|
-
block_max = max(block_max, s);
|
|
189
|
-
@@ -202,8 +186,6 @@
|
|
190
|
-
}
|
|
191
|
-
}
|
|
192
|
-
|
|
193
|
-
- // Accumulate V contribution by tiling head_dim again.
|
|
194
|
-
- // Barriers must be uniform, so only math is guarded by valid_query.
|
|
195
|
-
for (var ht: u32 = 0u; ht < num_head_tiles; ht = ht + 1u) {
|
|
196
|
-
let d0 = ht * HEAD_TILE;
|
|
197
|
-
let tile_len = min(HEAD_TILE, head_dim - d0);
|
|
198
|
-
@@ -217,7 +199,7 @@
|
|
199
|
-
}
|
|
200
|
-
} else {
|
|
201
|
-
for (var td: u32 = 0u; td < tile_len; td = td + 1u) {
|
|
202
|
-
- shared_V[thread_idx * HEAD_TILE + td] = 0.0;
|
|
203
|
-
+ shared_V[thread_idx * HEAD_TILE + td] = f16(0.0);
|
|
204
|
-
}
|
|
205
|
-
}
|
|
206
|
-
|
|
207
|
-
@@ -230,7 +212,6 @@
|
|
208
|
-
if (is_masked(query_pos, key_pos)) { continue; }
|
|
209
|
-
|
|
210
|
-
let p = exp(scores[k] - m_new);
|
|
211
|
-
- // Only accumulate l_i on first head tile to avoid double counting
|
|
212
|
-
if (ht == 0u) {
|
|
213
|
-
l_i = l_i + p;
|
|
214
|
-
}
|
|
215
|
-
@@ -249,10 +230,9 @@
|
|
216
|
-
}
|
|
217
|
-
}
|
|
218
|
-
|
|
219
|
-
- // Always write output when valid_query - write zeros if l_i is 0 to avoid garbage
|
|
220
|
-
if (valid_query) {
|
|
221
|
-
let out_offset = query_pos * num_heads * head_dim + head_idx * head_dim;
|
|
222
|
-
- let inv_l_i = select(0.0, 1.0 / l_i, l_i > 0.0);
|
|
223
|
-
+ let inv_l_i = select(f16(0.0), f16(1.0) / l_i, l_i > f16(0.0));
|
|
224
|
-
for (var d: u32 = 0u; d < head_dim; d = d + 1u) {
|
|
225
|
-
output[out_offset + d] = acc[d] * inv_l_i;
|
|
226
|
-
}
|