doppler-gpu 0.2.2 → 0.4.3
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 +182 -0
- package/README.md +20 -8
- package/assets/doppler.svg +1601 -0
- package/package.json +82 -33
- package/src/cli/cli-model-resolution.js +275 -0
- package/src/cli/cli-output.js +453 -0
- package/src/cli/doppler-cli.js +817 -709
- package/src/cli/doppler-quickstart.d.ts +7 -23
- package/src/cli/doppler-quickstart.js +20 -4
- package/src/cli/doppler-serve.js +1 -1
- package/src/cli/runtime-profiles.d.ts +40 -0
- package/src/cli/runtime-profiles.js +225 -0
- package/src/client/doppler-api.js +1 -6
- package/src/client/doppler-registry.d.ts +12 -0
- package/src/client/doppler-registry.js +27 -0
- package/src/client/doppler-registry.json +79 -5
- package/src/client/failure-taxonomy.js +139 -0
- package/src/client/fault-injection.js +30 -0
- package/src/client/provider.d.ts +151 -0
- package/src/client/provider.js +374 -0
- package/src/client/receipt.js +156 -0
- package/src/client/runtime/index.js +56 -4
- package/src/client/runtime/model-manager.d.ts +0 -14
- package/src/client/runtime/model-manager.js +2 -42
- package/src/client/runtime/model-session.js +33 -0
- package/src/client/runtime/model-source.d.ts +13 -1
- package/src/client/runtime/model-source.js +128 -2
- package/src/client/runtime/node-quickstart-cache.d.ts +22 -0
- package/src/client/runtime/node-quickstart-cache.js +353 -0
- package/src/client/runtime/source-runtime.js +1 -10
- package/src/client/runtime/types.d.ts +0 -60
- package/src/client/wrap-pipeline-handle.d.ts +20 -0
- package/src/client/wrap-pipeline-handle.js +84 -0
- package/src/config/README.md +3 -1
- package/src/config/conversion/README.md +23 -157
- package/src/config/conversion/embeddinggemma/google-embeddinggemma-300m-q4k-ehf16-af32.json +19 -4
- package/src/config/conversion/gemma3/gemma-3-1b-it-f16-af32.json +16 -9
- package/src/config/conversion/gemma3/gemma-3-1b-it-q4k-ehf16-af32.json +24 -9
- package/src/config/conversion/gemma3/gemma-3-270m-it-q4k-ehf16-af32.json +24 -9
- package/src/config/conversion/gemma3/translategemma-4b-1b-enes-q4k-ehf16-af32.json +16 -9
- package/src/config/conversion/gemma3/translategemma-4b-it-q4k-ehf16-af32.json +24 -9
- package/src/config/conversion/gemma4/gemma-4-31b-it-text-q4k-ehf16-af16.json +719 -0
- package/src/config/conversion/gemma4/gemma-4-31b-it-text-q4k-ehf16-af32.json +715 -0
- package/src/config/conversion/gemma4/gemma-4-e2b-it-q4k-ehf16-af16-int4ple.json +589 -0
- package/src/config/conversion/gemma4/gemma-4-e2b-it-q4k-ehf16-af32-int4ple.json +566 -0
- package/src/config/conversion/gemma4/gemma-4-e2b-it-q4k-ehf16-af32.json +368 -62
- package/src/config/conversion/gemma4/gemma-4-moe-q4k-ehf16-af32.json +212 -55
- package/src/config/conversion/gpt-oss-20b-f16-xmxfp4.json +216 -57
- package/src/config/conversion/janus/janus-pro-1b-text-q4k-ehaf16.json +229 -57
- package/src/config/conversion/lfm2/lfm2.5-1.2b-instruct-q4k-ehf16-af32.json +17 -10
- package/src/config/conversion/qwen3/qwen-3-5-0-8b-q4k-ehaf16.json +75 -33
- package/src/config/conversion/qwen3/qwen-3-5-2b-q4k-ehaf16.json +91 -59
- package/src/config/conversion/qwen3/qwen-3-6-27b-q4k-eaf16.json +517 -0
- package/src/config/conversion/qwen3/qwen-3-6-27b-q4k-ehaf16.json +511 -0
- package/src/config/execution-contract-check.js +6 -5
- package/src/config/kernel-path-loader.d.ts +8 -2
- package/src/config/kernels/kernel-ref-digests.js +125 -122
- package/src/config/kernels/registry.d.ts +0 -35
- package/src/config/kernels/registry.js +5 -9
- package/src/config/kernels/registry.json +615 -368
- package/src/config/merge-helpers.d.ts +1 -1
- package/src/config/merge-helpers.js +1 -1
- package/src/config/merge.js +35 -7
- package/src/config/param-categories.d.ts +0 -9
- package/src/config/param-categories.js +1 -5
- package/src/config/param-validator.d.ts +2 -0
- package/src/config/param-validator.js +16 -0
- package/src/config/platforms/loader.d.ts +0 -27
- package/src/config/platforms/loader.js +0 -30
- package/src/config/required-inference-fields-contract-check.js +1 -7
- package/src/config/runtime/README.md +28 -0
- package/src/config/runtime/experiments/bench/gemma4-bench-q4k.json +32 -0
- package/src/config/runtime/experiments/bench/qwen3-bench-q4k-batched.json +2 -2
- package/src/config/runtime/experiments/debug/gemma4-e2b-browser-diagnose-q4k.json +85 -0
- package/src/config/runtime/platform/metal-apple-q4k.json +19 -3
- package/src/config/runtime/profiles/gemma4-31b-f16-activations-probe.json +65 -0
- package/src/config/runtime/profiles/gemma4-e2b-baseline-probe.json +41 -0
- package/src/config/runtime/profiles/gemma4-e2b-fused-rmsnorm-probe.json +44 -0
- package/src/config/runtime/profiles/gemma4-e2b-fused-rmsnorm-test.json +19 -0
- package/src/config/runtime/profiles/gemma4-e2b-int4ple-baseline-probe.json +31 -0
- package/src/config/runtime/profiles/gemma4-e2b-int4ple-f16a-probe.json +34 -0
- package/src/config/runtime/profiles/gemma4-e2b-int4ple-reference-export.json +64 -0
- package/src/config/runtime/profiles/gemma4-e2b-int4ple-retain-q4k-probe.json +47 -0
- package/src/config/runtime/profiles/gemma4-e2b-prefill-profile.json +35 -0
- package/src/config/runtime/profiles/gemma4-e2b-throughput-probe.json +16 -0
- package/src/config/runtime/profiles/qwen-3-5-0-8b-prefill-profile.json +35 -0
- package/src/config/runtime/profiles/qwen-3-5-0-8b-throughput-probe.json +16 -0
- package/src/config/runtime/profiles/qwen3-6-27b-f16-activations-probe.json +62 -0
- package/src/config/schema/benchmark.schema.d.ts +2 -0
- package/src/config/schema/benchmark.schema.js +1 -0
- package/src/config/schema/bridge.schema.d.ts +16 -0
- package/src/config/schema/browser-suite-metrics.schema.d.ts +1 -0
- package/src/config/schema/browser-suite-metrics.schema.js +2 -0
- package/src/config/schema/converter.schema.d.ts +5 -0
- package/src/config/schema/converter.schema.js +3 -0
- package/src/config/schema/debug.schema.d.ts +22 -1
- package/src/config/schema/debug.schema.js +1 -0
- package/src/config/schema/distill-training.schema.d.ts +0 -1
- package/src/config/schema/distill-training.schema.js +1 -1
- package/src/config/schema/doppler.schema.d.ts +21 -3
- package/src/config/schema/doppler.schema.js +116 -9
- package/src/config/schema/emulation.schema.d.ts +3 -0
- package/src/config/schema/emulation.schema.js +16 -16
- package/src/config/schema/execution-v1.schema.d.ts +12 -1
- package/src/config/schema/execution-v1.schema.js +7 -1
- package/src/config/schema/index.d.ts +54 -1
- package/src/config/schema/index.js +15 -0
- package/src/config/schema/inference.schema.d.ts +3 -3
- package/src/config/schema/kernel-path.schema.d.ts +6 -27
- package/src/config/schema/kernel-thresholds.schema.d.ts +0 -2
- package/src/config/schema/kernel-thresholds.schema.js +3 -7
- package/src/config/schema/kvcache.schema.js +0 -9
- package/src/config/schema/loading.schema.d.ts +12 -0
- package/src/config/schema/loading.schema.js +4 -0
- package/src/config/schema/manifest.schema.d.ts +73 -1
- package/src/config/schema/manifest.schema.js +48 -2
- package/src/config/schema/moe.schema.js +0 -8
- package/src/config/schema/program-bundle.schema.d.ts +233 -0
- package/src/config/schema/program-bundle.schema.js +393 -0
- package/src/config/schema/ul-training.schema.d.ts +0 -1
- package/src/config/schema/ul-training.schema.js +1 -1
- package/src/config/source-packages/litert/gemma-4-e2b-it.json +36 -15
- package/src/config/supported-operations.d.ts +14 -0
- package/src/config/supported-operations.js +34 -0
- package/src/config/supported-operations.json +178 -0
- package/src/config/training-defaults.js +1 -7
- package/src/config/transforms/capability-transform-resolver.js +3 -1
- package/src/config/transforms/execution-graph-transforms.d.ts +55 -0
- package/src/config/transforms/execution-graph-transforms.js +693 -29
- package/src/converter/conversion-plan.js +13 -10
- package/src/converter/core.d.ts +57 -0
- package/src/converter/core.js +582 -13
- package/src/converter/parsers/diffusion.js +1 -33
- package/src/converter/quantization-info.d.ts +5 -4
- package/src/converter/quantization-info.js +39 -3
- package/src/converter/quantizer.d.ts +19 -0
- package/src/converter/quantizer.js +81 -4
- package/src/debug/config.d.ts +0 -4
- package/src/debug/config.js +0 -17
- package/src/debug/index.d.ts +28 -30
- package/src/debug/index.js +0 -18
- package/src/debug/stats.d.ts +0 -8
- package/src/debug/stats.js +3 -3
- package/src/debug/tensor.d.ts +9 -2
- package/src/debug/tensor.js +11 -4
- package/src/debug/trace.js +2 -15
- package/src/errors/doppler-error.js +20 -0
- package/src/experimental/browser/quantization.d.ts +5 -0
- package/src/experimental/distribution/shard-delivery.js +179 -23
- package/src/experimental/training/artifacts.js +1 -14
- package/src/experimental/training/checkpoint.js +1 -14
- package/src/experimental/training/datasets/token-batch.d.ts +1 -1
- package/src/experimental/training/datasets/token-batch.js +2 -2
- package/src/experimental/training/distillation/dataset.js +2 -15
- package/src/experimental/training/distillation/eval.js +3 -3
- package/src/experimental/training/distillation/runtime.js +3 -16
- package/src/experimental/training/distillation/student-fixture.js +6 -6
- package/src/experimental/training/lora-pipeline.js +1 -14
- package/src/experimental/training/objectives/base.d.ts +2 -2
- package/src/experimental/training/objectives/cross_entropy.d.ts +2 -2
- package/src/experimental/training/objectives/distill_kd.d.ts +2 -2
- package/src/experimental/training/objectives/distill_kd.js +2 -2
- package/src/experimental/training/objectives/distill_triplet.d.ts +2 -2
- package/src/experimental/training/objectives/distill_triplet.js +2 -2
- package/src/experimental/training/objectives/ul_stage1_joint.d.ts +2 -2
- package/src/experimental/training/objectives/ul_stage2_base.d.ts +2 -2
- package/src/experimental/training/objectives/ul_stage2_base.js +1 -1
- package/src/experimental/training/operator-artifacts.js +1 -14
- package/src/experimental/training/runner.js +1 -14
- package/src/experimental/training/workload-packs/registry.json +6 -6
- package/src/experimental/training/workloads.js +1 -1
- package/src/formats/gguf/types.d.ts +0 -63
- package/src/formats/gguf/types.js +58 -38
- package/src/formats/litert/types.d.ts +0 -13
- package/src/formats/litert/types.js +2 -2
- package/src/formats/rdrr/distributed/index.d.ts +34 -0
- package/src/formats/rdrr/distributed/index.js +3 -0
- package/src/formats/rdrr/distributed/parsing.d.ts +6 -0
- package/src/formats/rdrr/distributed/parsing.js +23 -0
- package/src/formats/rdrr/distributed/validation.d.ts +23 -0
- package/src/formats/rdrr/distributed/validation.js +131 -0
- package/src/formats/rdrr/groups.d.ts +1 -13
- package/src/formats/rdrr/groups.js +0 -34
- package/src/formats/rdrr/index.d.ts +9 -0
- package/src/formats/rdrr/index.js +9 -0
- package/src/formats/rdrr/integrity.d.ts +51 -0
- package/src/formats/rdrr/integrity.js +200 -0
- package/src/formats/rdrr/lowerings.d.ts +28 -0
- package/src/formats/rdrr/lowerings.js +112 -0
- package/src/formats/rdrr/manifest.d.ts +0 -14
- package/src/formats/rdrr/manifest.js +5 -58
- package/src/formats/rdrr/merkle.d.ts +24 -0
- package/src/formats/rdrr/merkle.js +118 -0
- package/src/formats/rdrr/parsing.js +4 -0
- package/src/formats/rdrr/slice-compiler.d.ts +29 -0
- package/src/formats/rdrr/slice-compiler.js +188 -0
- package/src/formats/rdrr/source-transform-contract.js +390 -0
- package/src/formats/rdrr/storage-descriptor.d.ts +36 -0
- package/src/formats/rdrr/storage-descriptor.js +186 -0
- package/src/formats/rdrr/tensor-config-validator.d.ts +0 -9
- package/src/formats/rdrr/tensor-config-validator.js +0 -30
- package/src/formats/rdrr/types.d.ts +115 -1
- package/src/formats/rdrr/validation.js +245 -0
- package/src/formats/tflite/types.d.ts +8 -31
- package/src/formats/tflite/types.js +50 -5
- package/src/gpu/command-recorder.d.ts +9 -21
- package/src/gpu/command-recorder.js +41 -2
- package/src/gpu/device.d.ts +17 -3
- package/src/gpu/device.js +20 -4
- package/src/gpu/kernel-runtime.d.ts +8 -15
- package/src/gpu/kernel-selection-log.d.ts +0 -8
- package/src/gpu/kernel-selection-log.js +0 -15
- package/src/gpu/kernel-tuner/cache.js +0 -3
- package/src/gpu/kernels/README.md +50 -18
- package/src/gpu/kernels/activation-static-qdq.d.ts +22 -0
- package/src/gpu/kernels/activation-static-qdq.js +69 -0
- package/src/gpu/kernels/activation_static_qdq.wgsl +25 -0
- package/src/gpu/kernels/attention.d.ts +28 -19
- package/src/gpu/kernels/attention.js +443 -9
- package/src/gpu/kernels/attention_decode_contiguous_turboquant_f16kv.wgsl +0 -1
- package/src/gpu/kernels/attention_decode_contiguous_turboquant_prod_f16kv.wgsl +0 -1
- package/src/gpu/kernels/attention_decode_f16.wgsl +1 -1
- package/src/gpu/kernels/attention_decode_f16kv.wgsl +1 -1
- package/src/gpu/kernels/attention_decode_online_f16.wgsl +32 -29
- package/src/gpu/kernels/attention_decode_online_f16kv.wgsl +1 -1
- package/src/gpu/kernels/attention_decode_paged_f16kv.wgsl +1 -1
- package/src/gpu/kernels/attention_decode_tiered_f16kv.wgsl +1 -1
- package/src/gpu/kernels/attention_decode_tiered_turboquant_f16kv.wgsl +0 -1
- package/src/gpu/kernels/attention_decode_tiered_turboquant_prod_f16kv.wgsl +0 -1
- package/src/gpu/kernels/attention_f16.wgsl +88 -30
- package/src/gpu/kernels/attention_f16kv.wgsl +45 -15
- package/src/gpu/kernels/attention_head512_f16.wgsl +270 -0
- package/src/gpu/kernels/attention_head512_f16kv.wgsl +271 -0
- package/src/gpu/kernels/attention_prefill_flash_head256_f16kv.wgsl +306 -0
- package/src/gpu/kernels/attention_prefill_flash_reduce.wgsl +101 -0
- package/src/gpu/kernels/attention_small_f16.wgsl +22 -20
- package/src/gpu/kernels/attention_small_f16kv.wgsl +1 -1
- package/src/gpu/kernels/attention_streaming_f16.wgsl +39 -16
- package/src/gpu/kernels/attention_streaming_f16kv.wgsl +22 -3
- package/src/gpu/kernels/backward/index.d.ts +6 -0
- package/src/gpu/kernels/backward/layernorm_backward.wgsl +0 -1
- package/src/gpu/kernels/backward/utils.d.ts +35 -0
- package/src/gpu/kernels/bf16_to_f32.wgsl +0 -16
- package/src/gpu/kernels/bias_add_f16.wgsl +1 -1
- package/src/gpu/kernels/cast_f32_to_f16.wgsl +0 -1
- package/src/gpu/kernels/codegen/patches/attention_f16.from.attention.diff +19 -215
- package/src/gpu/kernels/codegen/patches/attention_f16kv.from.attention.diff +8 -82
- package/src/gpu/kernels/codegen/patches/attention_small_f16.from.attention_small.diff +92 -59
- package/src/gpu/kernels/codegen/patches/attention_streaming_f16.from.attention_streaming.diff +2 -85
- package/src/gpu/kernels/codegen/patches/gelu_f16.from.gelu.diff +27 -24
- package/src/gpu/kernels/codegen/patches/residual_f16.from.residual.diff +1 -1
- package/src/gpu/kernels/codegen/patches/rmsnorm_f16.from.rmsnorm.diff +249 -179
- package/src/gpu/kernels/codegen/patches/rope_f16.from.rope.diff +133 -235
- package/src/gpu/kernels/codegen/patches/sample_f16.from.sample.diff +119 -90
- package/src/gpu/kernels/codegen/wgsl-variants.js +142 -8
- package/src/gpu/kernels/constants.d.ts +9 -48
- package/src/gpu/kernels/constants.js +0 -23
- package/src/gpu/kernels/conv2d.d.ts +0 -4
- package/src/gpu/kernels/conv2d_f16.wgsl +1 -1
- package/src/gpu/kernels/dequant.d.ts +0 -16
- package/src/gpu/kernels/dequant.js +1 -71
- package/src/gpu/kernels/dequant_mxfp4_expert_f16.wgsl +1 -1
- package/src/gpu/kernels/dequant_shared.wgsl +1 -55
- package/src/gpu/kernels/dispatch.d.ts +12 -97
- package/src/gpu/kernels/dispatch.js +0 -117
- package/src/gpu/kernels/energy_eval_f16.wgsl +1 -1
- package/src/gpu/kernels/energy_quintel_grad_f16.wgsl +1 -1
- package/src/gpu/kernels/energy_quintel_reduce_f16.wgsl +1 -1
- package/src/gpu/kernels/energy_quintel_update_f16.wgsl +1 -1
- package/src/gpu/kernels/energy_update_f16.wgsl +1 -1
- package/src/gpu/kernels/fused-gate-up-gelu.js +121 -0
- package/src/gpu/kernels/fused_ffn_f16.wgsl +1 -1
- package/src/gpu/kernels/fused_ffn_q4k.wgsl +0 -19
- package/src/gpu/kernels/fused_ffn_q4k_f16.wgsl +0 -18
- package/src/gpu/kernels/fused_gate_up_gelu_gated_f16.wgsl +124 -0
- package/src/gpu/kernels/fused_matmul_q4.wgsl +23 -439
- package/src/gpu/kernels/fused_matmul_q4_batched_f16.wgsl +1 -1
- package/src/gpu/kernels/fused_matmul_q4_batched_f16a.wgsl +0 -1
- package/src/gpu/kernels/fused_matmul_q4_batched_f16acc_f16a.wgsl +158 -0
- package/src/gpu/kernels/fused_matmul_q4_f16a.wgsl +0 -1
- package/src/gpu/kernels/fused_matmul_q4_multicol_f16.wgsl +0 -1
- package/src/gpu/kernels/fused_matmul_q4_multicol_f16a.wgsl +32 -33
- package/src/gpu/kernels/fused_matmul_q4_multicol_f16a_f32acc.wgsl +209 -0
- package/src/gpu/kernels/fused_matmul_q4_prefill_tiled_f16.wgsl +261 -0
- package/src/gpu/kernels/fused_matmul_q4_widetile.wgsl +164 -0
- package/src/gpu/kernels/fused_matmul_q4_widetile_f16.wgsl +198 -0
- package/src/gpu/kernels/fused_matmul_q4_widetile_f16a.wgsl +161 -0
- package/src/gpu/kernels/fused_matmul_q4_widetile_residual.wgsl +173 -0
- package/src/gpu/kernels/fused_matmul_rmsnorm_f16.wgsl +1 -1
- package/src/gpu/kernels/fused_rmsnorm_q4_widetile.wgsl +271 -0
- package/src/gpu/kernels/fused_swiglu_f16.wgsl +1 -1
- package/src/gpu/kernels/gather.d.ts +0 -7
- package/src/gpu/kernels/gather_f16.wgsl +1 -1
- package/src/gpu/kernels/gelu_f16.wgsl +14 -15
- package/src/gpu/kernels/groupnorm.d.ts +0 -4
- package/src/gpu/kernels/groupnorm_apply_f16.wgsl +1 -1
- package/src/gpu/kernels/groupnorm_stats_f16.wgsl +1 -1
- package/src/gpu/kernels/index.d.ts +13 -5
- package/src/gpu/kernels/index.js +9 -4
- package/src/gpu/kernels/kv_quantize_turboquant.wgsl +0 -2
- package/src/gpu/kernels/kv_quantize_turboquant_prod.wgsl +0 -2
- package/src/gpu/kernels/layernorm.d.ts +0 -4
- package/src/gpu/kernels/layernorm_f16.wgsl +1 -1
- package/src/gpu/kernels/{sana_linear_attention.d.ts → linear_attention.d.ts} +5 -5
- package/src/gpu/kernels/{sana_linear_attention.js → linear_attention.js} +16 -16
- package/src/gpu/kernels/logit-merge.d.ts +0 -24
- package/src/gpu/kernels/logit-merge.js +1 -17
- package/src/gpu/kernels/matmul-dispatch.js +7 -1
- package/src/gpu/kernels/matmul-selection.js +68 -1
- package/src/gpu/kernels/matmul.d.ts +0 -28
- package/src/gpu/kernels/matmul.js +101 -7
- package/src/gpu/kernels/matmul_f16_tiled.wgsl +33 -34
- package/src/gpu/kernels/matmul_gemv_residual_f16.wgsl +1 -1
- package/src/gpu/kernels/matmul_gemv_subgroup.wgsl +0 -1
- package/src/gpu/kernels/matmul_gemv_subgroup_f16a.wgsl +24 -25
- package/src/gpu/kernels/modulate_f16.wgsl +1 -1
- package/src/gpu/kernels/moe.d.ts +0 -56
- package/src/gpu/kernels/moe.js +0 -15
- package/src/gpu/kernels/moe_gather_f16.wgsl +1 -1
- package/src/gpu/kernels/pixel_shuffle_f16.wgsl +1 -1
- package/src/gpu/kernels/residual.d.ts +0 -12
- package/src/gpu/kernels/residual_f16.wgsl +2 -2
- package/src/gpu/kernels/residual_f16_vec4.wgsl +5 -5
- package/src/gpu/kernels/rmsnorm.d.ts +0 -8
- package/src/gpu/kernels/rmsnorm.js +17 -1
- package/src/gpu/kernels/rmsnorm.wgsl +5 -6
- package/src/gpu/kernels/rmsnorm_f16.wgsl +65 -39
- package/src/gpu/kernels/rope.d.ts +1 -5
- package/src/gpu/kernels/rope.js +12 -2
- package/src/gpu/kernels/rope.wgsl +33 -19
- package/src/gpu/kernels/rope_f16.wgsl +39 -25
- package/src/gpu/kernels/sample.d.ts +0 -5
- package/src/gpu/kernels/sample.wgsl +7 -6
- package/src/gpu/kernels/sample_f16.wgsl +28 -30
- package/src/gpu/kernels/scatter_add_dynamic_f16.wgsl +1 -1
- package/src/gpu/kernels/shader-cache.d.ts +13 -0
- package/src/gpu/kernels/shader-cache.js +27 -0
- package/src/gpu/kernels/silu_f16.wgsl +1 -1
- package/src/gpu/kernels/softmax.wgsl +4 -144
- package/src/gpu/kernels/split_qg_f16.wgsl +1 -1
- package/src/gpu/kernels/split_qkv_f16.wgsl +1 -1
- package/src/gpu/kernels/topk_f16.wgsl +1 -1
- package/src/gpu/kernels/turboquant-codebook.js +11 -40
- package/src/gpu/kernels/uniform-utils.d.ts +18 -0
- package/src/gpu/kernels/upsample2d.d.ts +0 -4
- package/src/gpu/kernels/upsample2d_f16.wgsl +1 -1
- package/src/gpu/kernels/utils.d.ts +28 -0
- package/src/gpu/kernels/utils.js +41 -0
- package/src/gpu/perf-guards.d.ts +0 -12
- package/src/gpu/perf-guards.js +0 -62
- package/src/gpu/profiler.d.ts +0 -21
- package/src/gpu/profiler.js +0 -34
- package/src/gpu/submit-tracker.d.ts +0 -57
- package/src/gpu/submit-tracker.js +3 -62
- package/src/gpu/tensor.d.ts +0 -23
- package/src/gpu/tensor.js +0 -41
- package/src/gpu/uniform-cache.d.ts +0 -11
- package/src/gpu/uniform-cache.js +0 -23
- package/src/index.d.ts +1 -0
- package/src/index.js +1 -0
- package/src/inference/browser-harness-model-helpers.d.ts +0 -9
- package/src/inference/browser-harness-model-helpers.js +22 -13
- package/src/inference/browser-harness-runtime-helpers.d.ts +0 -5
- package/src/inference/browser-harness-runtime-helpers.js +2 -2
- package/src/inference/browser-harness-suite-helpers.d.ts +4 -1
- package/src/inference/browser-harness-suite-helpers.js +87 -1
- package/src/inference/browser-harness-text-helpers.d.ts +13 -0
- package/src/inference/browser-harness-text-helpers.js +267 -18
- package/src/inference/browser-harness.d.ts +0 -29
- package/src/inference/browser-harness.js +271 -71
- package/src/inference/kv-cache/base.d.ts +6 -0
- package/src/inference/kv-cache/base.js +16 -1
- package/src/inference/kv-cache/mixed-geometry.d.ts +1 -0
- package/src/inference/kv-cache/mixed-geometry.js +12 -2
- package/src/inference/kv-cache/quantized.js +2 -0
- package/src/inference/kv-cache/tiered.js +7 -0
- package/src/inference/kv-cache/types.d.ts +2 -0
- package/src/inference/multi-model-network.d.ts +1 -1
- package/src/inference/multi-model-network.js +1 -2
- package/src/inference/network-evolution.d.ts +0 -12
- package/src/inference/network-evolution.js +2 -2
- package/src/inference/pipelines/context.js +21 -2
- package/src/inference/pipelines/diffusion/init.js +2 -2
- package/src/inference/pipelines/diffusion/pipeline.js +41 -96
- package/src/inference/pipelines/diffusion/text-encoder-gpu.d.ts +0 -5
- package/src/inference/pipelines/diffusion/text-encoder-gpu.js +0 -283
- package/src/inference/pipelines/diffusion/vae.js +4 -4
- package/src/inference/pipelines/energy/quintel.d.ts +15 -0
- package/src/inference/pipelines/structured/index.d.ts +2 -0
- package/src/inference/pipelines/structured/index.js +2 -0
- package/src/inference/pipelines/structured/json-grammar-mask.js +143 -0
- package/src/inference/pipelines/structured/tsv-grammar-mask.js +166 -0
- package/src/inference/pipelines/text/attention/attn-config.d.ts +0 -5
- package/src/inference/pipelines/text/attention/attn-config.js +1 -1
- package/src/inference/pipelines/text/attention/projections.d.ts +2 -0
- package/src/inference/pipelines/text/attention/projections.js +70 -23
- package/src/inference/pipelines/text/attention/record.js +118 -5
- package/src/inference/pipelines/text/attention/rmsnorm-fusion-gate.d.ts +8 -0
- package/src/inference/pipelines/text/attention/rmsnorm-fusion-gate.js +19 -0
- package/src/inference/pipelines/text/attention/run.js +97 -6
- package/src/inference/pipelines/text/attention/types.d.ts +2 -0
- package/src/inference/pipelines/text/attention/types.js +9 -4
- package/src/inference/pipelines/text/chat-format.d.ts +0 -10
- package/src/inference/pipelines/text/chat-format.js +1 -8
- package/src/inference/pipelines/text/config.d.ts +3 -25
- package/src/inference/pipelines/text/config.js +105 -3
- package/src/inference/pipelines/text/debug-utils/utils.js +2 -23
- package/src/inference/pipelines/text/drift-policy.js +1 -1
- package/src/inference/pipelines/text/embed.d.ts +23 -27
- package/src/inference/pipelines/text/embed.js +10 -43
- package/src/inference/pipelines/text/execution-plan.d.ts +5 -3
- package/src/inference/pipelines/text/execution-plan.js +15 -2
- package/src/inference/pipelines/text/execution-runtime-builders.d.ts +0 -4
- package/src/inference/pipelines/text/execution-runtime-builders.js +0 -11
- package/src/inference/pipelines/text/execution-v1.d.ts +18 -0
- package/src/inference/pipelines/text/execution-v1.js +94 -25
- package/src/inference/pipelines/text/ffn/dense.js +180 -3
- package/src/inference/pipelines/text/ffn/standard.js +40 -13
- package/src/inference/pipelines/text/generator-decode-policy.js +243 -0
- package/src/inference/pipelines/text/generator-helpers.d.ts +6 -2
- package/src/inference/pipelines/text/generator-helpers.js +20 -2
- package/src/inference/pipelines/text/generator-prefill-helpers.js +293 -0
- package/src/inference/pipelines/text/generator-runtime.d.ts +0 -7
- package/src/inference/pipelines/text/generator-runtime.js +8 -6
- package/src/inference/pipelines/text/generator-steps.d.ts +22 -7
- package/src/inference/pipelines/text/generator-steps.js +23 -9
- package/src/inference/pipelines/text/generator.d.ts +0 -6
- package/src/inference/pipelines/text/generator.js +254 -530
- package/src/inference/pipelines/text/init-chat-templates.js +86 -0
- package/src/inference/pipelines/text/init-qkv-fusion.js +201 -0
- package/src/inference/pipelines/text/init-rope.js +408 -0
- package/src/inference/pipelines/text/init.d.ts +8 -2
- package/src/inference/pipelines/text/init.js +190 -710
- package/src/inference/pipelines/text/kernel-trace.d.ts +0 -33
- package/src/inference/pipelines/text/kernel-trace.js +0 -33
- package/src/inference/pipelines/text/layer-plan-gpu.js +542 -0
- package/src/inference/pipelines/text/layer-plan.d.ts +0 -5
- package/src/inference/pipelines/text/layer-plan.js +1 -1
- package/src/inference/pipelines/text/layer.d.ts +70 -8
- package/src/inference/pipelines/text/layer.js +105 -537
- package/src/inference/pipelines/text/linear-attention.d.ts +2 -1
- package/src/inference/pipelines/text/load-errors.d.ts +21 -0
- package/src/inference/pipelines/text/load-errors.js +59 -0
- package/src/inference/pipelines/text/logits/cpu.js +2 -20
- package/src/inference/pipelines/text/logits/gpu.d.ts +5 -4
- package/src/inference/pipelines/text/logits/gpu.js +148 -11
- package/src/inference/pipelines/text/logits/index.d.ts +1 -1
- package/src/inference/pipelines/text/model-load.d.ts +5 -4
- package/src/inference/pipelines/text/model-load.js +74 -17
- package/src/inference/pipelines/text/ops.d.ts +13 -2
- package/src/inference/pipelines/text/ops.js +1 -1
- package/src/inference/pipelines/text/per-layer-inputs.js +347 -59
- package/src/inference/pipelines/text/probes.d.ts +3 -6
- package/src/inference/pipelines/text/probes.js +198 -31
- package/src/inference/pipelines/text/sampling.d.ts +0 -5
- package/src/inference/pipelines/text/sampling.js +140 -0
- package/src/inference/pipelines/text/stage-names.js +12 -0
- package/src/inference/pipelines/text/state.js +2 -0
- package/src/inference/pipelines/text/tsir-fixture-writer.js +409 -0
- package/src/inference/pipelines/text/types.d.ts +8 -78
- package/src/inference/pipelines/text/weights.d.ts +0 -50
- package/src/inference/pipelines/text/weights.js +2 -51
- package/src/inference/pipelines/text.d.ts +9 -1
- package/src/inference/pipelines/text.js +233 -29
- package/src/inference/pipelines/vision/index.js +0 -107
- package/src/inference/runtime-model.d.ts +0 -7
- package/src/inference/runtime-model.js +0 -23
- package/src/inference/test-harness.d.ts +6 -22
- package/src/inference/test-harness.js +97 -23
- package/src/inference/tokenizer.js +1 -10
- package/src/inference/tokenizers/sentencepiece.js +65 -15
- package/src/inference/tokenizers/types.d.ts +0 -21
- package/src/loader/doppler-loader.d.ts +3 -0
- package/src/loader/doppler-loader.js +90 -12
- package/src/loader/embedding-loader.js +30 -5
- package/src/loader/loader-types.d.ts +4 -0
- package/src/loader/manifest-config.d.ts +2 -7
- package/src/loader/manifest-config.js +25 -1
- package/src/loader/memory-monitor.d.ts +17 -5
- package/src/loader/per-layer-input-loader.js +114 -1
- package/src/loader/quantization-constants.d.ts +6 -0
- package/src/loader/shard-resolver.d.ts +1 -0
- package/src/loader/shard-resolver.js +15 -0
- package/src/loader/tensors/source-transform.d.ts +15 -0
- package/src/loader/tensors/source-transform.js +317 -34
- package/src/loader/tensors/tensor-loader.d.ts +2 -9
- package/src/loader/tensors/tensor-loader.js +84 -1
- package/src/loader/tensors/tensor-reader.js +28 -2
- package/src/memory/address-table.d.ts +1 -9
- package/src/memory/buffer-pool.d.ts +0 -16
- package/src/memory/buffer-pool.js +1 -33
- package/src/memory/heap-manager.d.ts +18 -43
- package/src/memory/unified-detect.d.ts +19 -34
- package/src/models/embeddinggemma.d.ts +7 -0
- package/src/models/embeddinggemma.js +24 -0
- package/src/models/family.d.ts +24 -0
- package/src/models/family.js +25 -0
- package/src/models/gemma3.d.ts +7 -0
- package/src/models/gemma3.js +33 -0
- package/src/models/gemma4.d.ts +7 -0
- package/src/models/gemma4.js +33 -0
- package/src/models/qwen3.d.ts +7 -0
- package/src/models/qwen3.js +36 -0
- package/src/rules/converter/tokenizer.rules.json +5 -0
- package/src/rules/execution-rules-contract-check.js +72 -3
- package/src/rules/inference/capability-transforms.rules.json +87 -0
- package/src/rules/inference/execution.rules.json +11 -1
- package/src/rules/inference/ffn.rules.json +6 -2
- package/src/rules/loader/tensor-loader.rules.json +9 -0
- package/src/rules/rule-registry.js +53 -75
- package/src/storage/artifact-storage-context.d.ts +12 -3
- package/src/storage/artifact-storage-context.js +389 -98
- package/src/storage/backends/opfs-store.js +138 -88
- package/src/storage/blake3.js +9 -1
- package/src/storage/export.js +1 -5
- package/src/storage/normalize-model-id.js +3 -0
- package/src/storage/preflight.js +22 -0
- package/src/storage/quickstart-downloader.js +44 -56
- package/src/storage/reports.js +35 -11
- package/src/storage/shard-manager.d.ts +8 -2
- package/src/storage/shard-manager.js +176 -15
- package/src/storage/source-artifact-store.d.ts +0 -2
- package/src/storage/source-artifact-store.js +3 -9
- package/src/sw.js +0 -3
- package/src/tooling/browser-command-runner.js +39 -0
- package/src/tooling/command-api-constants.js +1 -1
- package/src/tooling/command-api-family-normalizers.d.ts +1 -0
- package/src/tooling/command-api-family-normalizers.js +49 -7
- package/src/tooling/command-api-helpers.d.ts +0 -2
- package/src/tooling/command-api-helpers.js +6 -2
- package/src/tooling/command-api.d.ts +20 -1
- package/src/tooling/command-api.js +10 -1
- package/src/tooling/command-envelope.d.ts +0 -1
- package/src/tooling/command-envelope.js +1 -1
- package/src/tooling/command-runner-shared.d.ts +10 -0
- package/src/tooling/conversion-config-materializer.d.ts +0 -10
- package/src/tooling/conversion-config-materializer.js +1 -1
- package/src/tooling/diagnose-runner.js +1 -4
- package/src/tooling/hf-registry-utils.d.ts +2 -6
- package/src/tooling/hf-registry-utils.js +90 -19
- package/src/tooling/kernel-path-builder/index.js +1 -4
- package/src/tooling/kernel-path-builder/runtime-overlay.js +1 -3
- package/src/tooling/lean-execution-contract-runner.d.ts +0 -10
- package/src/tooling/lean-execution-contract-runner.js +2 -2
- package/src/tooling/litert-package-runtime.d.ts +1 -1
- package/src/tooling/litert-package-runtime.js +373 -21
- package/src/tooling/node-browser-command-runner.d.ts +15 -0
- package/src/tooling/node-browser-command-runner.js +97 -13
- package/src/tooling/node-command-runner.js +36 -0
- package/src/tooling/node-convert-worker-pool.js +12 -0
- package/src/tooling/node-convert-worker.js +10 -1
- package/src/tooling/node-converter.js +90 -1
- package/src/tooling/node-source-runtime.js +257 -150
- package/src/tooling/opfs-cache.js +1 -4
- package/src/tooling/precision-replay-math.js +174 -0
- package/src/tooling/program-bundle-host.d.ts +16 -0
- package/src/tooling/program-bundle-host.js +20 -0
- package/src/tooling/program-bundle-parity.d.ts +28 -0
- package/src/tooling/program-bundle-parity.js +210 -0
- package/src/tooling/program-bundle.d.ts +61 -0
- package/src/tooling/program-bundle.js +1119 -0
- package/src/tooling/rdrr-integrity-refresh.d.ts +32 -0
- package/src/tooling/rdrr-integrity-refresh.js +271 -0
- package/src/tooling/reference-verify.js +191 -0
- package/src/tooling/source-artifact-adapter.d.ts +0 -20
- package/src/tooling/source-artifact-adapter.js +4 -4
- package/src/tooling/source-package-profiles.js +1 -7
- package/src/tooling/source-runtime-bundle.d.ts +27 -0
- package/src/tooling/source-runtime-bundle.js +56 -23
- package/src/tooling/source-runtime-converter-config.d.ts +0 -5
- package/src/tooling/source-runtime-converter-config.js +13 -9
- package/src/tooling/source-runtime-materializer.js +4 -8
- package/src/tooling-exports/device.d.ts +11 -0
- package/src/tooling-exports/device.js +14 -0
- package/src/tooling-exports/manifest.d.ts +8 -0
- package/src/tooling-exports/manifest.js +11 -0
- package/src/tooling-exports/storage.d.ts +25 -0
- package/src/tooling-exports/storage.js +29 -0
- package/src/tooling-exports/structured.d.ts +1 -0
- package/src/tooling-exports/structured.js +5 -0
- package/src/tooling-exports.d.ts +23 -0
- package/src/tooling-exports.js +17 -0
- package/src/tooling-exports.shared.d.ts +27 -30
- package/src/tooling-exports.shared.js +28 -29
- package/src/types/index.d.ts +0 -3
- package/src/types/index.js +0 -3
- package/src/utils/array-buffer.d.ts +4 -0
- package/src/utils/array-buffer.js +9 -0
- package/src/utils/canonical-hash.d.ts +7 -0
- package/src/utils/canonical-hash.js +74 -0
- package/src/utils/clone-json.d.ts +1 -0
- package/src/utils/clone-json.js +7 -0
- package/src/utils/encode-utf8.js +5 -0
- package/src/utils/index.d.ts +4 -1
- package/src/utils/index.js +4 -1
- package/src/utils/runtime-env.js +7 -0
- package/src/utils/sha256.d.ts +1 -0
- package/src/utils/sha256.js +18 -4
- package/src/utils/stable-sort-object.d.ts +1 -0
- package/src/utils/stable-sort-object.js +13 -0
- package/src/version.js +1 -1
- package/tools/convert-safetensors-node.js +1 -2
- package/src/bootstrap.d.ts +0 -1
- package/src/config/conversion/sana/sana-sprint-0.6b-f16.json +0 -247
- package/src/config/kernels/kernel-ref.d.ts +0 -17
- package/src/config/kernels/kernel-ref.js +0 -75
- package/src/config/runtime/kernels/dequant-f16-q4k.json +0 -13
- package/src/config/runtime/kernels/dequant-f32-q4k.json +0 -13
- package/src/config/runtime/kernels/embeddinggemma-q4k-dequant-f32a.json +0 -37
- package/src/config/runtime/kernels/fused-q4k.json +0 -13
- package/src/config/runtime/kernels/gemma2-q4k-dequant-f16a.json +0 -33
- package/src/config/runtime/kernels/gemma2-q4k-dequant-f32a-nosubgroups.json +0 -33
- package/src/config/runtime/kernels/gemma2-q4k-fused-f32a.json +0 -33
- package/src/config/runtime/kernels/safe-q4k.json +0 -13
- package/src/config/runtime/overlays/debug-f32a.json +0 -3
- package/src/config/runtime/overlays/trace-config.json +0 -28
- package/src/config/runtime/profiles/gemma4-e2b-throughput.json +0 -39
- package/src/debug/perf.d.ts +0 -33
- package/src/debug/perf.js +0 -51
- package/src/errors/index.d.ts +0 -1
- package/src/errors/index.js +0 -1
- package/src/formats/gguf/index.d.ts +0 -8
- package/src/formats/gguf/index.js +0 -4
- package/src/formats/litert/index.d.ts +0 -1
- package/src/formats/litert/index.js +0 -1
- package/src/formats/safetensors/index.d.ts +0 -8
- package/src/formats/safetensors/index.js +0 -4
- package/src/formats/tflite/index.d.ts +0 -1
- package/src/formats/tflite/index.js +0 -1
- package/src/gpu/kernels/attention_decode_optimized.wgsl +0 -445
- package/src/gpu/kernels/dequant_mxfp4_vec4.wgsl +0 -116
- package/src/gpu/kernels/dequant_q8_0.wgsl +0 -98
- package/src/gpu/kernels/dequant_shared_vec4.wgsl +0 -160
- package/src/gpu/kernels/moe_gather_vec4.wgsl +0 -74
- package/src/inference/functiongemma.d.ts +0 -15
- package/src/inference/functiongemma.js +0 -1
- package/src/inference/pipelines/diffusion/sana-transformer.d.ts +0 -53
- package/src/inference/pipelines/diffusion/sana-transformer.js +0 -738
- package/src/rules/kernels/kernel-validator.d.ts +0 -24
- package/src/rules/kernels/kernel-validator.js +0 -160
- package/src/sw.d.ts +0 -1
- package/src/tooling/bench-runner.js +0 -64
- package/src/tooling/distill-runner.js +0 -56
- package/src/tooling/lora-runner.js +0 -37
- package/src/tooling/verify-runner.js +0 -63
- package/src/types/gpu.d.ts +0 -185
- package/src/types/gpu.js +0 -5
- package/src/types/inference.d.ts +0 -197
- package/src/types/inference.js +0 -5
- package/src/types/model.d.ts +0 -130
- package/src/types/model.js +0 -5
- package/src/utils/format-bytes.js +0 -20
- /package/src/config/{runtime/diagnostics → diagnostics}/drift-policies.json +0 -0
- /package/src/gpu/kernels/{sana_linear_attention_apply.wgsl → linear_attention_apply.wgsl} +0 -0
- /package/src/gpu/kernels/{sana_linear_attention_apply_f16.wgsl → linear_attention_apply_f16.wgsl} +0 -0
- /package/src/gpu/kernels/{sana_linear_attention_summary.wgsl → linear_attention_summary.wgsl} +0 -0
- /package/src/gpu/kernels/{sana_linear_attention_summary_f16.wgsl → linear_attention_summary_f16.wgsl} +0 -0
- /package/src/rules/kernels/{sana-linear-attention.rules.json → linear-attention.rules.json} +0 -0
|
@@ -8,9 +8,18 @@ import {
|
|
|
8
8
|
isWeightBuffer,
|
|
9
9
|
} from '../../../gpu/weight-buffer.js';
|
|
10
10
|
import { createTensor } from '../../../gpu/tensor.js';
|
|
11
|
-
import {
|
|
11
|
+
import {
|
|
12
|
+
castF16ToF32,
|
|
13
|
+
castF32ToF16,
|
|
14
|
+
recordActivationStaticQdq,
|
|
15
|
+
recordCastF16ToF32,
|
|
16
|
+
recordCastF32ToF16,
|
|
17
|
+
recordScale,
|
|
18
|
+
runActivationStaticQdq,
|
|
19
|
+
runScale,
|
|
20
|
+
} from '../../../gpu/kernel-selector.js';
|
|
12
21
|
import { getNormWeightBuffer, getWeightBuffer } from './weights.js';
|
|
13
|
-
import { doMatmul, doRMSNorm, releaseOrTrack } from './ops.js';
|
|
22
|
+
import { doCast, doMatmul, doRMSNorm, releaseOrTrack } from './ops.js';
|
|
14
23
|
import { runProbes } from './probes.js';
|
|
15
24
|
import { embed, isRangeBackedCpuEmbeddingSource, normalizeRangeBytes, decodeRangeChunkIntoOutput } from './embed.js';
|
|
16
25
|
import { selectRuleValue } from '../../../rules/rule-registry.js';
|
|
@@ -19,6 +28,8 @@ import { acquireBuffer, releaseBuffer } from '../../../memory/buffer-pool.js';
|
|
|
19
28
|
|
|
20
29
|
const pleRuntimeCache = new WeakMap();
|
|
21
30
|
const pleRangeRowCache = new WeakMap();
|
|
31
|
+
const PLE_ACTIVATION_STATIC_QMIN = -128;
|
|
32
|
+
const PLE_ACTIVATION_STATIC_QMAX = 127;
|
|
22
33
|
|
|
23
34
|
function getPerLayerInputWeights(context) {
|
|
24
35
|
const weights = context.weights.get('per_layer_inputs');
|
|
@@ -63,6 +74,106 @@ function isDensePleProjectionDtype(dtype) {
|
|
|
63
74
|
return dtype === 'f16' || dtype === 'f32';
|
|
64
75
|
}
|
|
65
76
|
|
|
77
|
+
function resolvePleActivationStaticScaleValue(value, label) {
|
|
78
|
+
if (value == null) {
|
|
79
|
+
return null;
|
|
80
|
+
}
|
|
81
|
+
const parsed = Number(value);
|
|
82
|
+
if (!Number.isFinite(parsed) || parsed <= 0) {
|
|
83
|
+
throw new Error(
|
|
84
|
+
`Gemma 4 per-layer projection activation static scale "${label}" must be > 0. ` +
|
|
85
|
+
`Got "${String(value)}".`
|
|
86
|
+
);
|
|
87
|
+
}
|
|
88
|
+
return parsed;
|
|
89
|
+
}
|
|
90
|
+
|
|
91
|
+
function resolvePleActivationStaticQdqConfig(perLayerInputWeights) {
|
|
92
|
+
const inputScale = resolvePleActivationStaticScaleValue(
|
|
93
|
+
perLayerInputWeights?.perLayerModelProjectionInputActivationStaticScale,
|
|
94
|
+
'input'
|
|
95
|
+
);
|
|
96
|
+
const outputScale = resolvePleActivationStaticScaleValue(
|
|
97
|
+
perLayerInputWeights?.perLayerModelProjectionOutputActivationStaticScale,
|
|
98
|
+
'output'
|
|
99
|
+
);
|
|
100
|
+
if (inputScale == null && outputScale == null) {
|
|
101
|
+
return null;
|
|
102
|
+
}
|
|
103
|
+
if (inputScale == null || outputScale == null) {
|
|
104
|
+
throw new Error(
|
|
105
|
+
'Gemma 4 per-layer projection static activation quantization requires both input and output scales.'
|
|
106
|
+
);
|
|
107
|
+
}
|
|
108
|
+
return {
|
|
109
|
+
inputScale,
|
|
110
|
+
outputScale,
|
|
111
|
+
};
|
|
112
|
+
}
|
|
113
|
+
|
|
114
|
+
function getTensorElementCount(tensor, label) {
|
|
115
|
+
if (!Array.isArray(tensor?.shape) || tensor.shape.length === 0) {
|
|
116
|
+
throw new Error(`${label} requires a tensor with an explicit shape.`);
|
|
117
|
+
}
|
|
118
|
+
const count = tensor.shape.reduce((total, dimension) => total * Number(dimension), 1);
|
|
119
|
+
if (!Number.isFinite(count) || count <= 0) {
|
|
120
|
+
throw new Error(`${label} resolved an invalid tensor element count (${String(count)}).`);
|
|
121
|
+
}
|
|
122
|
+
return count;
|
|
123
|
+
}
|
|
124
|
+
|
|
125
|
+
async function applyPleActivationStaticQdq(tensor, scale, recorder, decodeBuffers, label) {
|
|
126
|
+
const count = getTensorElementCount(tensor, label);
|
|
127
|
+
if (tensor.dtype === 'f32') {
|
|
128
|
+
return recorder
|
|
129
|
+
? await recordActivationStaticQdq(recorder, tensor, scale, {
|
|
130
|
+
count,
|
|
131
|
+
qmin: PLE_ACTIVATION_STATIC_QMIN,
|
|
132
|
+
qmax: PLE_ACTIVATION_STATIC_QMAX,
|
|
133
|
+
})
|
|
134
|
+
: await runActivationStaticQdq(tensor, scale, {
|
|
135
|
+
count,
|
|
136
|
+
qmin: PLE_ACTIVATION_STATIC_QMIN,
|
|
137
|
+
qmax: PLE_ACTIVATION_STATIC_QMAX,
|
|
138
|
+
});
|
|
139
|
+
}
|
|
140
|
+
if (tensor.dtype !== 'f16') {
|
|
141
|
+
throw new Error(
|
|
142
|
+
`${label} requires f16/f32 activations for static activation quantize/dequantize. ` +
|
|
143
|
+
`Got "${String(tensor.dtype)}".`
|
|
144
|
+
);
|
|
145
|
+
}
|
|
146
|
+
|
|
147
|
+
let widenedTensor = null;
|
|
148
|
+
let qdqTensor = null;
|
|
149
|
+
try {
|
|
150
|
+
widenedTensor = recorder
|
|
151
|
+
? await recordCastF16ToF32(recorder, tensor)
|
|
152
|
+
: await castF16ToF32(tensor);
|
|
153
|
+
qdqTensor = recorder
|
|
154
|
+
? await recordActivationStaticQdq(recorder, widenedTensor, scale, {
|
|
155
|
+
count,
|
|
156
|
+
qmin: PLE_ACTIVATION_STATIC_QMIN,
|
|
157
|
+
qmax: PLE_ACTIVATION_STATIC_QMAX,
|
|
158
|
+
})
|
|
159
|
+
: await runActivationStaticQdq(widenedTensor, scale, {
|
|
160
|
+
count,
|
|
161
|
+
qmin: PLE_ACTIVATION_STATIC_QMIN,
|
|
162
|
+
qmax: PLE_ACTIVATION_STATIC_QMAX,
|
|
163
|
+
});
|
|
164
|
+
return recorder
|
|
165
|
+
? await recordCastF32ToF16(recorder, qdqTensor)
|
|
166
|
+
: await castF32ToF16(qdqTensor);
|
|
167
|
+
} finally {
|
|
168
|
+
if (qdqTensor) {
|
|
169
|
+
releaseOrTrack(recorder, qdqTensor.buffer, decodeBuffers);
|
|
170
|
+
}
|
|
171
|
+
if (widenedTensor) {
|
|
172
|
+
releaseOrTrack(recorder, widenedTensor.buffer, decodeBuffers);
|
|
173
|
+
}
|
|
174
|
+
}
|
|
175
|
+
}
|
|
176
|
+
|
|
66
177
|
export function resolveDensePleProjectionWeight(
|
|
67
178
|
weight,
|
|
68
179
|
label = 'per_layer_model_projection'
|
|
@@ -1229,6 +1340,8 @@ export async function preparePerLayerInputs(tokenIds, inputEmbedsTensor, context
|
|
|
1229
1340
|
if (projectionDevice == null) {
|
|
1230
1341
|
throw new Error('Gemma 4 per-layer input projection requires an initialized GPU device.');
|
|
1231
1342
|
}
|
|
1343
|
+
const pleActivationStaticQdq = resolvePleActivationStaticQdqConfig(perLayerInputWeights);
|
|
1344
|
+
let projectionInputTensor = inputEmbedsTensor;
|
|
1232
1345
|
|
|
1233
1346
|
// Decode-path optimizations: coalesced PLE read + fused projection matmul.
|
|
1234
1347
|
// Gated on numTokens === 1 (decode) and row-major embeddings (non-transpose).
|
|
@@ -1237,8 +1350,12 @@ export async function preparePerLayerInputs(tokenIds, inputEmbedsTensor, context
|
|
|
1237
1350
|
const embedTranspose = (hasSplitEmbeddingTables || useHotVocabularyTables) ? false : getEmbeddingTranspose(embedTokensPerLayer);
|
|
1238
1351
|
const canFuseDecodeOps = numTokens === 1 && !embedTranspose && !hasRangeBackedProjection;
|
|
1239
1352
|
const tokenIdsAreGpuBuffer = isGpuBufferInstance(tokenIds);
|
|
1240
|
-
|
|
1241
|
-
|
|
1353
|
+
|
|
1354
|
+
const tokenIdHint = Number.isInteger(options.tokenIdHint) ? Number(options.tokenIdHint) : null;
|
|
1355
|
+
const decodeTokenId = canFuseDecodeOps
|
|
1356
|
+
? (tokenIdsAreGpuBuffer
|
|
1357
|
+
? (Number.isInteger(tokenIdHint) ? tokenIdHint : null)
|
|
1358
|
+
: Number(tokenIds[0]))
|
|
1242
1359
|
: null;
|
|
1243
1360
|
|
|
1244
1361
|
if (decodeTokenId != null) {
|
|
@@ -1310,55 +1427,110 @@ export async function preparePerLayerInputs(tokenIds, inputEmbedsTensor, context
|
|
|
1310
1427
|
// Produces [numTokens × totalPerLayerHiddenSize], then scales the full output and
|
|
1311
1428
|
// extracts per-layer slices via GPU buffer copies.
|
|
1312
1429
|
let fusedProjectionSlices = null;
|
|
1313
|
-
if (canFuseDecodeOps) {
|
|
1314
|
-
const fusedProjection = await doMatmul(
|
|
1315
|
-
inputEmbedsTensor,
|
|
1316
|
-
projectionWeight,
|
|
1317
|
-
numTokens,
|
|
1318
|
-
totalPerLayerHiddenSize,
|
|
1319
|
-
config.hiddenSize,
|
|
1320
|
-
{
|
|
1321
|
-
transposeB: 'auto',
|
|
1322
|
-
label: 'per_layer_fused_projection',
|
|
1323
|
-
kernelPath: context.kernelPath ?? null,
|
|
1324
|
-
role: 'per_layer_model_projection',
|
|
1325
|
-
outputDtype: activationDtype,
|
|
1326
|
-
},
|
|
1327
|
-
recorder
|
|
1328
|
-
);
|
|
1329
|
-
|
|
1330
|
-
// One scale dispatch instead of numLayers
|
|
1331
|
-
const scaledFused = recorder
|
|
1332
|
-
? await recordScale(recorder, fusedProjection, projectionScale, {
|
|
1333
|
-
count: numTokens * totalPerLayerHiddenSize,
|
|
1334
|
-
})
|
|
1335
|
-
: await runScale(fusedProjection, projectionScale, {
|
|
1336
|
-
count: numTokens * totalPerLayerHiddenSize,
|
|
1337
|
-
});
|
|
1338
|
-
releaseOrTrack(recorder, fusedProjection.buffer, decodeBuffers);
|
|
1339
|
-
|
|
1340
|
-
// Step 4: Extract per-layer slices via GPU buffer copies (one encoder, one submit).
|
|
1341
|
-
// Reuse cached slice buffers when available to avoid per-step pool churn.
|
|
1342
|
-
const device = getDevice();
|
|
1343
|
-
const sliceBytes = hiddenSizePerLayerInput * activationBytesPerElement;
|
|
1344
|
-
const encoder = recorder ? recorder.getEncoder() : device.createCommandEncoder();
|
|
1345
|
-
fusedProjectionSlices = new Array(numLayers);
|
|
1346
|
-
for (let l = 0; l < numLayers; l++) {
|
|
1347
|
-
const sliceBuf = pleCache?.sliceBuffers?.[l] ?? acquireBuffer(sliceBytes, undefined, `L${l}.per_layer_proj_slice`);
|
|
1348
|
-
encoder.copyBufferToBuffer(scaledFused.buffer, l * sliceBytes, sliceBuf, 0, sliceBytes);
|
|
1349
|
-
fusedProjectionSlices[l] = sliceBuf;
|
|
1350
|
-
}
|
|
1351
|
-
if (!recorder) {
|
|
1352
|
-
device.queue.submit([encoder.finish()]);
|
|
1353
|
-
}
|
|
1354
|
-
releaseOrTrack(recorder, scaledFused.buffer, decodeBuffers);
|
|
1355
|
-
}
|
|
1356
|
-
|
|
1357
1430
|
const embedDtypeResolved = selectRuleValue('inference', 'dtype', 'f16OrF32FromDtype', {
|
|
1358
1431
|
dtype: perLayerEmbeddingDtype,
|
|
1359
1432
|
});
|
|
1360
1433
|
|
|
1361
1434
|
try {
|
|
1435
|
+
if (pleActivationStaticQdq) {
|
|
1436
|
+
projectionInputTensor = await applyPleActivationStaticQdq(
|
|
1437
|
+
inputEmbedsTensor,
|
|
1438
|
+
pleActivationStaticQdq.inputScale,
|
|
1439
|
+
recorder,
|
|
1440
|
+
decodeBuffers,
|
|
1441
|
+
'Gemma 4 per-layer projection input'
|
|
1442
|
+
);
|
|
1443
|
+
}
|
|
1444
|
+
|
|
1445
|
+
if (canFuseDecodeOps) {
|
|
1446
|
+
let fusedProjection = null;
|
|
1447
|
+
let fusedProjectionForScale = null;
|
|
1448
|
+
let scaledFused = null;
|
|
1449
|
+
try {
|
|
1450
|
+
// Per-layer-model projection: precision island. Source rows decoded
|
|
1451
|
+
// from int4_per_row produce activations whose magnitude before
|
|
1452
|
+
// RMSNorm-mediated scale-down can exceed f16 max (~65504). When the
|
|
1453
|
+
// matmul kernel does not internally widen its accumulator (RDNA3 is
|
|
1454
|
+
// strict, Metal happens to widen), the f16 output saturates to Inf
|
|
1455
|
+
// and then NaN. Compute matmul + QDQ + scale in f32 so the dynamic
|
|
1456
|
+
// range fits, then narrow to activationDtype after the scale step
|
|
1457
|
+
// brings magnitudes back to ~O(1).
|
|
1458
|
+
fusedProjection = await doMatmul(
|
|
1459
|
+
projectionInputTensor,
|
|
1460
|
+
projectionWeight,
|
|
1461
|
+
numTokens,
|
|
1462
|
+
totalPerLayerHiddenSize,
|
|
1463
|
+
config.hiddenSize,
|
|
1464
|
+
{
|
|
1465
|
+
transposeB: 'auto',
|
|
1466
|
+
label: 'per_layer_fused_projection',
|
|
1467
|
+
kernelPath: context.kernelPath ?? null,
|
|
1468
|
+
role: 'per_layer_model_projection',
|
|
1469
|
+
outputDtype: 'f32',
|
|
1470
|
+
},
|
|
1471
|
+
recorder
|
|
1472
|
+
);
|
|
1473
|
+
fusedProjectionForScale = pleActivationStaticQdq
|
|
1474
|
+
? await applyPleActivationStaticQdq(
|
|
1475
|
+
fusedProjection,
|
|
1476
|
+
pleActivationStaticQdq.outputScale,
|
|
1477
|
+
recorder,
|
|
1478
|
+
decodeBuffers,
|
|
1479
|
+
'Gemma 4 per-layer projection output'
|
|
1480
|
+
)
|
|
1481
|
+
: fusedProjection;
|
|
1482
|
+
if (fusedProjectionForScale !== fusedProjection) {
|
|
1483
|
+
releaseOrTrack(recorder, fusedProjection.buffer, decodeBuffers);
|
|
1484
|
+
fusedProjection = null;
|
|
1485
|
+
}
|
|
1486
|
+
|
|
1487
|
+
let scaledFusedF32 = recorder
|
|
1488
|
+
? await recordScale(recorder, fusedProjectionForScale, projectionScale, {
|
|
1489
|
+
count: numTokens * totalPerLayerHiddenSize,
|
|
1490
|
+
})
|
|
1491
|
+
: await runScale(fusedProjectionForScale, projectionScale, {
|
|
1492
|
+
count: numTokens * totalPerLayerHiddenSize,
|
|
1493
|
+
});
|
|
1494
|
+
releaseOrTrack(recorder, fusedProjectionForScale.buffer, decodeBuffers);
|
|
1495
|
+
fusedProjectionForScale = null;
|
|
1496
|
+
|
|
1497
|
+
scaledFused = scaledFusedF32.dtype === activationDtype
|
|
1498
|
+
? scaledFusedF32
|
|
1499
|
+
: await doCast(scaledFusedF32, activationDtype, recorder);
|
|
1500
|
+
if (scaledFused !== scaledFusedF32) {
|
|
1501
|
+
releaseOrTrack(recorder, scaledFusedF32.buffer, decodeBuffers);
|
|
1502
|
+
}
|
|
1503
|
+
|
|
1504
|
+
// Step 4: Extract per-layer slices via GPU buffer copies (one encoder, one submit).
|
|
1505
|
+
// Reuse cached slice buffers when available to avoid per-step pool churn.
|
|
1506
|
+
const device = getDevice();
|
|
1507
|
+
const sliceBytes = hiddenSizePerLayerInput * activationBytesPerElement;
|
|
1508
|
+
const encoder = recorder ? recorder.getEncoder() : device.createCommandEncoder();
|
|
1509
|
+
fusedProjectionSlices = new Array(numLayers);
|
|
1510
|
+
for (let l = 0; l < numLayers; l++) {
|
|
1511
|
+
const sliceBuf = pleCache?.sliceBuffers?.[l] ?? acquireBuffer(sliceBytes, undefined, `L${l}.per_layer_proj_slice`);
|
|
1512
|
+
encoder.copyBufferToBuffer(scaledFused.buffer, l * sliceBytes, sliceBuf, 0, sliceBytes);
|
|
1513
|
+
fusedProjectionSlices[l] = sliceBuf;
|
|
1514
|
+
}
|
|
1515
|
+
if (!recorder) {
|
|
1516
|
+
device.queue.submit([encoder.finish()]);
|
|
1517
|
+
}
|
|
1518
|
+
releaseOrTrack(recorder, scaledFused.buffer, decodeBuffers);
|
|
1519
|
+
scaledFused = null;
|
|
1520
|
+
} catch (error) {
|
|
1521
|
+
if (scaledFused) {
|
|
1522
|
+
releaseOrTrack(recorder, scaledFused.buffer, decodeBuffers);
|
|
1523
|
+
}
|
|
1524
|
+
if (fusedProjectionForScale && fusedProjectionForScale !== fusedProjection) {
|
|
1525
|
+
releaseOrTrack(recorder, fusedProjectionForScale.buffer, decodeBuffers);
|
|
1526
|
+
}
|
|
1527
|
+
if (fusedProjection) {
|
|
1528
|
+
releaseOrTrack(recorder, fusedProjection.buffer, decodeBuffers);
|
|
1529
|
+
}
|
|
1530
|
+
throw error;
|
|
1531
|
+
}
|
|
1532
|
+
}
|
|
1533
|
+
|
|
1362
1534
|
for (let layerIdx = 0; layerIdx < numLayers; layerIdx++) {
|
|
1363
1535
|
const hiddenOffset = layerIdx * hiddenSizePerLayerInput;
|
|
1364
1536
|
const layerEmbedSource = useHotVocabularyTables
|
|
@@ -1374,6 +1546,7 @@ export async function preparePerLayerInputs(tokenIds, inputEmbedsTensor, context
|
|
|
1374
1546
|
hiddenSize: hiddenSizePerLayerInput,
|
|
1375
1547
|
vocabSize: useHotVocabularyTables ? (hotVocabularyRuntime.sentinelIndex + 1) : vocabSizePerLayerInput,
|
|
1376
1548
|
scaleEmbeddings: true,
|
|
1549
|
+
probeStage: 'per_layer_embed_out',
|
|
1377
1550
|
recorder,
|
|
1378
1551
|
numTokens,
|
|
1379
1552
|
indexOffset: perLayerTokenIds ? perLayerIndexOffset : indexOffset,
|
|
@@ -1390,6 +1563,7 @@ export async function preparePerLayerInputs(tokenIds, inputEmbedsTensor, context
|
|
|
1390
1563
|
outputBuffer: canFuseDecodeOps
|
|
1391
1564
|
? (pleCache?.gatherSliceBuffers?.[layerIdx] ?? undefined)
|
|
1392
1565
|
: undefined,
|
|
1566
|
+
stats: context.stats ?? null,
|
|
1393
1567
|
});
|
|
1394
1568
|
|
|
1395
1569
|
if (fusedProjectionSlices) {
|
|
@@ -1406,6 +1580,7 @@ export async function preparePerLayerInputs(tokenIds, inputEmbedsTensor, context
|
|
|
1406
1580
|
let projectionWeightBufferForLayer = null;
|
|
1407
1581
|
let projectionWeightOffset = 0;
|
|
1408
1582
|
let projectedTensor = null;
|
|
1583
|
+
let projectionTensorForScale = null;
|
|
1409
1584
|
try {
|
|
1410
1585
|
const rangeBackedProjection = await loadRangeBackedPleProjectionSliceBytes(
|
|
1411
1586
|
perLayerModelProjection,
|
|
@@ -1448,8 +1623,12 @@ export async function preparePerLayerInputs(tokenIds, inputEmbedsTensor, context
|
|
|
1448
1623
|
});
|
|
1449
1624
|
projectionWeightOffset = hiddenOffset * config.hiddenSize * projectionWeightBytes;
|
|
1450
1625
|
}
|
|
1626
|
+
// Per-layer-model projection: precision island (see fused path
|
|
1627
|
+
// above for rationale). Force f32 through matmul + QDQ + scale,
|
|
1628
|
+
// then narrow to activationDtype after scale brings magnitudes
|
|
1629
|
+
// back to ~O(1).
|
|
1451
1630
|
projectedTensor = await doMatmul(
|
|
1452
|
-
|
|
1631
|
+
projectionInputTensor,
|
|
1453
1632
|
projectionWeightForLayer,
|
|
1454
1633
|
numTokens,
|
|
1455
1634
|
hiddenSizePerLayerInput,
|
|
@@ -1461,24 +1640,70 @@ export async function preparePerLayerInputs(tokenIds, inputEmbedsTensor, context
|
|
|
1461
1640
|
layerIdx,
|
|
1462
1641
|
kernelPath: context.kernelPath ?? null,
|
|
1463
1642
|
role: 'per_layer_model_projection',
|
|
1464
|
-
outputDtype:
|
|
1643
|
+
outputDtype: 'f32',
|
|
1465
1644
|
},
|
|
1466
1645
|
recorder
|
|
1467
1646
|
);
|
|
1647
|
+
await runProbes('per_layer_projection_in', projectedTensor.buffer, {
|
|
1648
|
+
layerIdx,
|
|
1649
|
+
numTokens,
|
|
1650
|
+
hiddenSize: hiddenSizePerLayerInput,
|
|
1651
|
+
probes: context.debugProbes,
|
|
1652
|
+
recorder,
|
|
1653
|
+
operatorDiagnostics: context.operatorDiagnostics,
|
|
1654
|
+
dtype: 'f32',
|
|
1655
|
+
});
|
|
1656
|
+
projectionTensorForScale = pleActivationStaticQdq
|
|
1657
|
+
? await applyPleActivationStaticQdq(
|
|
1658
|
+
projectedTensor,
|
|
1659
|
+
pleActivationStaticQdq.outputScale,
|
|
1660
|
+
recorder,
|
|
1661
|
+
decodeBuffers,
|
|
1662
|
+
`Gemma 4 per-layer projection output L${layerIdx}`
|
|
1663
|
+
)
|
|
1664
|
+
: projectedTensor;
|
|
1665
|
+
if (projectionTensorForScale !== projectedTensor) {
|
|
1666
|
+
releaseOrTrack(recorder, projectedTensor.buffer, decodeBuffers);
|
|
1667
|
+
projectedTensor = null;
|
|
1668
|
+
}
|
|
1669
|
+
const scaledF32 = recorder
|
|
1670
|
+
? await recordScale(recorder, projectionTensorForScale, projectionScale, {
|
|
1671
|
+
count: numTokens * hiddenSizePerLayerInput,
|
|
1672
|
+
})
|
|
1673
|
+
: await runScale(projectionTensorForScale, projectionScale, {
|
|
1674
|
+
count: numTokens * hiddenSizePerLayerInput,
|
|
1675
|
+
});
|
|
1676
|
+
if (projectionTensorForScale === projectedTensor) {
|
|
1677
|
+
projectedTensor = null;
|
|
1678
|
+
}
|
|
1679
|
+
releaseOrTrack(recorder, projectionTensorForScale.buffer, decodeBuffers);
|
|
1680
|
+
projectionTensorForScale = null;
|
|
1681
|
+
scaledProjectionTensor = scaledF32.dtype === activationDtype
|
|
1682
|
+
? scaledF32
|
|
1683
|
+
: await doCast(scaledF32, activationDtype, recorder);
|
|
1684
|
+
if (scaledProjectionTensor !== scaledF32) {
|
|
1685
|
+
releaseOrTrack(recorder, scaledF32.buffer, decodeBuffers);
|
|
1686
|
+
}
|
|
1468
1687
|
} finally {
|
|
1688
|
+
if (projectionTensorForScale && projectionTensorForScale !== projectedTensor) {
|
|
1689
|
+
releaseOrTrack(recorder, projectionTensorForScale.buffer, decodeBuffers);
|
|
1690
|
+
}
|
|
1691
|
+
if (projectedTensor) {
|
|
1692
|
+
releaseOrTrack(recorder, projectedTensor.buffer, decodeBuffers);
|
|
1693
|
+
}
|
|
1469
1694
|
if (projectionWeightBufferForLayer) {
|
|
1470
1695
|
releaseOrTrack(recorder, projectionWeightBufferForLayer, decodeBuffers);
|
|
1471
1696
|
}
|
|
1472
1697
|
}
|
|
1473
|
-
scaledProjectionTensor
|
|
1474
|
-
|
|
1475
|
-
|
|
1476
|
-
|
|
1477
|
-
:
|
|
1478
|
-
|
|
1479
|
-
|
|
1480
|
-
|
|
1481
|
-
|
|
1698
|
+
await runProbes('per_layer_projection_scaled', scaledProjectionTensor.buffer, {
|
|
1699
|
+
layerIdx,
|
|
1700
|
+
numTokens,
|
|
1701
|
+
hiddenSize: hiddenSizePerLayerInput,
|
|
1702
|
+
probes: context.debugProbes,
|
|
1703
|
+
recorder,
|
|
1704
|
+
operatorDiagnostics: context.operatorDiagnostics,
|
|
1705
|
+
dtype: activationDtype,
|
|
1706
|
+
});
|
|
1482
1707
|
}
|
|
1483
1708
|
|
|
1484
1709
|
// Fuse the residual add into RMSNorm so Gemma 4 PLE decode avoids an
|
|
@@ -1552,6 +1777,9 @@ export async function preparePerLayerInputs(tokenIds, inputEmbedsTensor, context
|
|
|
1552
1777
|
}
|
|
1553
1778
|
throw error;
|
|
1554
1779
|
} finally {
|
|
1780
|
+
if (projectionInputTensor !== inputEmbedsTensor) {
|
|
1781
|
+
releaseOrTrack(recorder, projectionInputTensor.buffer, decodeBuffers);
|
|
1782
|
+
}
|
|
1555
1783
|
if (!usesCachedScaledProjectionNormWeight && !isGpuBufferInstance(perLayerProjectionNorm)) {
|
|
1556
1784
|
releaseOrTrack(recorder, projectionNormWeight, decodeBuffers);
|
|
1557
1785
|
}
|
|
@@ -1568,6 +1796,66 @@ export async function preparePerLayerInputs(tokenIds, inputEmbedsTensor, context
|
|
|
1568
1796
|
);
|
|
1569
1797
|
}
|
|
1570
1798
|
|
|
1799
|
+
// Prefill hot-cache seeding: when we computed a multi-token batch and pleCache is
|
|
1800
|
+
// provided, extract per-token row slices and store unique token IDs in the cache.
|
|
1801
|
+
// Copies are recorded via the active recorder (if any) so they execute after the
|
|
1802
|
+
// batch computation and before the first decode step reads from the cache.
|
|
1803
|
+
// Without a recorder, the batch buffers are already populated so copies are safe
|
|
1804
|
+
// to submit immediately.
|
|
1805
|
+
if (pleCache) {
|
|
1806
|
+
const plePolicy = getPleHotCachePolicy(context.perLayerInputsSession ?? null);
|
|
1807
|
+
if (plePolicy && plePolicy.mode === 'prepared_tokens'
|
|
1808
|
+
&& pleCache.preparedTokenEntries instanceof Map
|
|
1809
|
+
&& !tokenIdsAreGpuBuffer
|
|
1810
|
+
&& numTokens > 0
|
|
1811
|
+
) {
|
|
1812
|
+
const tokenIdArray = Array.isArray(tokenIds) ? tokenIds : Array.from(tokenIds);
|
|
1813
|
+
const sliceBytes = hiddenSizePerLayerInput * activationBytesPerElement;
|
|
1814
|
+
const seen = new Set();
|
|
1815
|
+
const device = getDevice();
|
|
1816
|
+
let pendingEncoder = null;
|
|
1817
|
+
for (let tokenPos = 0; tokenPos < tokenIdArray.length; tokenPos++) {
|
|
1818
|
+
const tid = tokenIdArray[tokenPos];
|
|
1819
|
+
if (seen.has(tid) || pleCache.preparedTokenEntries.has(tid)) continue;
|
|
1820
|
+
if (pleCache.preparedTokenEntries.size >= plePolicy.maxTokens) break;
|
|
1821
|
+
seen.add(tid);
|
|
1822
|
+
const srcOffset = tokenPos * sliceBytes;
|
|
1823
|
+
const sliceBuffers = new Array(numLayers).fill(null);
|
|
1824
|
+
let allValid = true;
|
|
1825
|
+
for (let layerIdx = 0; layerIdx < numLayers; layerIdx++) {
|
|
1826
|
+
const srcBuffer = perLayerBuffers[layerIdx];
|
|
1827
|
+
if (!srcBuffer || srcBuffer.size < srcOffset + sliceBytes) {
|
|
1828
|
+
allValid = false;
|
|
1829
|
+
break;
|
|
1830
|
+
}
|
|
1831
|
+
const dstBuffer = acquireBuffer(sliceBytes, undefined, `ple_seed_L${layerIdx}`);
|
|
1832
|
+
if (recorder) {
|
|
1833
|
+
recorder.getEncoder().copyBufferToBuffer(srcBuffer, srcOffset, dstBuffer, 0, sliceBytes);
|
|
1834
|
+
} else {
|
|
1835
|
+
if (!pendingEncoder) pendingEncoder = device.createCommandEncoder();
|
|
1836
|
+
pendingEncoder.copyBufferToBuffer(srcBuffer, srcOffset, dstBuffer, 0, sliceBytes);
|
|
1837
|
+
}
|
|
1838
|
+
sliceBuffers[layerIdx] = dstBuffer;
|
|
1839
|
+
}
|
|
1840
|
+
if (!allValid) {
|
|
1841
|
+
for (const buf of sliceBuffers) {
|
|
1842
|
+
if (buf) releaseBuffer(buf);
|
|
1843
|
+
}
|
|
1844
|
+
continue;
|
|
1845
|
+
}
|
|
1846
|
+
storePreparedTokenEntry(
|
|
1847
|
+
pleCache, tid, sliceBuffers,
|
|
1848
|
+
context.perLayerInputsSession ?? null,
|
|
1849
|
+
activationDtype,
|
|
1850
|
+
context.stats ?? null
|
|
1851
|
+
);
|
|
1852
|
+
}
|
|
1853
|
+
if (pendingEncoder && device) {
|
|
1854
|
+
device.queue.submit([pendingEncoder.finish()]);
|
|
1855
|
+
}
|
|
1856
|
+
}
|
|
1857
|
+
}
|
|
1858
|
+
|
|
1571
1859
|
return perLayerBuffers;
|
|
1572
1860
|
}
|
|
1573
1861
|
|
|
@@ -20,12 +20,9 @@ export function runProbes(
|
|
|
20
20
|
hiddenSize: number;
|
|
21
21
|
probes?: ProbeConfigSchema[] | null;
|
|
22
22
|
recorder?: CommandRecorder | null;
|
|
23
|
+
operatorDiagnostics?: Record<string, unknown> | null;
|
|
23
24
|
dtype?: 'f16' | 'f32';
|
|
25
|
+
phase?: string | null;
|
|
26
|
+
tokenIndex?: number | null;
|
|
24
27
|
}
|
|
25
28
|
): Promise<void>;
|
|
26
|
-
|
|
27
|
-
export function hasProbeStage(
|
|
28
|
-
probes: ProbeConfigSchema[] | null | undefined,
|
|
29
|
-
stage: ProbeStage,
|
|
30
|
-
layerIdx?: number
|
|
31
|
-
): boolean;
|