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
|
@@ -0,0 +1,158 @@
|
|
|
1
|
+
// Fused Q4_K Matmul Kernel - W4A16 (F16 activations, F16 accum, F16 output)
|
|
2
|
+
//
|
|
3
|
+
// Batched prefill variant for the Gemma 4 31B experimental all-f16 lane.
|
|
4
|
+
// Unlike fused_matmul_q4_batched_f16a.wgsl, this kernel does not widen the
|
|
5
|
+
// reduction to f32. It keeps the same Q4_K block format and f16 output ABI, but
|
|
6
|
+
// distributes each output dot product across THREADS_PER_COL lanes at Q4_K
|
|
7
|
+
// subblock granularity so no single invocation serially accumulates more than
|
|
8
|
+
// 32 products into one f16 partial on Gemma 4 31B prefill shapes.
|
|
9
|
+
|
|
10
|
+
enable f16;
|
|
11
|
+
enable subgroups;
|
|
12
|
+
|
|
13
|
+
const QK_K: u32 = 256u;
|
|
14
|
+
const SUBBLOCK_SIZE: u32 = 32u;
|
|
15
|
+
|
|
16
|
+
struct Uniforms {
|
|
17
|
+
M: u32,
|
|
18
|
+
N: u32,
|
|
19
|
+
K: u32,
|
|
20
|
+
alpha: f32,
|
|
21
|
+
num_blocks_per_row: u32,
|
|
22
|
+
_pad0: u32,
|
|
23
|
+
_pad1: u32,
|
|
24
|
+
_pad2: u32,
|
|
25
|
+
}
|
|
26
|
+
|
|
27
|
+
struct Q4KBlock {
|
|
28
|
+
d_dmin: u32,
|
|
29
|
+
scales: array<u32, 3>,
|
|
30
|
+
qs: array<u32, 32>,
|
|
31
|
+
}
|
|
32
|
+
|
|
33
|
+
@group(0) @binding(0) var<uniform> u: Uniforms;
|
|
34
|
+
@group(0) @binding(1) var<storage, read> A: array<f16>;
|
|
35
|
+
@group(0) @binding(2) var<storage, read> B_q4k: array<Q4KBlock>;
|
|
36
|
+
@group(0) @binding(4) var<storage, read_write> C_f16: array<f16>;
|
|
37
|
+
|
|
38
|
+
fn unpack_f16_lo(packed: u32) -> f16 {
|
|
39
|
+
return f16(unpack2x16float(packed).x);
|
|
40
|
+
}
|
|
41
|
+
|
|
42
|
+
fn unpack_f16_hi(packed: u32) -> f16 {
|
|
43
|
+
return f16(unpack2x16float(packed).y);
|
|
44
|
+
}
|
|
45
|
+
|
|
46
|
+
fn get_scale_byte(scales: array<u32, 3>, byte_idx: u32) -> u32 {
|
|
47
|
+
let word_idx = byte_idx / 4u;
|
|
48
|
+
let byte_in_word = byte_idx % 4u;
|
|
49
|
+
return (scales[word_idx] >> (byte_in_word * 8u)) & 0xFFu;
|
|
50
|
+
}
|
|
51
|
+
|
|
52
|
+
fn get_scale_min_k4(scales: array<u32, 3>, j: u32) -> vec2<u32> {
|
|
53
|
+
var sc: u32;
|
|
54
|
+
var mn: u32;
|
|
55
|
+
|
|
56
|
+
if (j < 4u) {
|
|
57
|
+
sc = get_scale_byte(scales, j) & 63u;
|
|
58
|
+
mn = get_scale_byte(scales, j + 4u) & 63u;
|
|
59
|
+
} else {
|
|
60
|
+
let q_j = get_scale_byte(scales, j + 4u);
|
|
61
|
+
let q_lo = get_scale_byte(scales, j - 4u);
|
|
62
|
+
let q_hi = get_scale_byte(scales, j);
|
|
63
|
+
sc = (q_j & 0xFu) | ((q_lo >> 6u) << 4u);
|
|
64
|
+
mn = (q_j >> 4u) | ((q_hi >> 6u) << 4u);
|
|
65
|
+
}
|
|
66
|
+
return vec2<u32>(sc, mn);
|
|
67
|
+
}
|
|
68
|
+
|
|
69
|
+
fn get_q4(qs: array<u32, 32>, idx: u32) -> u32 {
|
|
70
|
+
let chunk = idx / 64u;
|
|
71
|
+
let pos_in_chunk = idx % 64u;
|
|
72
|
+
let use_upper = pos_in_chunk >= 32u;
|
|
73
|
+
let byte_in_range = select(pos_in_chunk, pos_in_chunk - 32u, use_upper);
|
|
74
|
+
let byte_idx = chunk * 32u + byte_in_range;
|
|
75
|
+
|
|
76
|
+
let word_idx = byte_idx / 4u;
|
|
77
|
+
let byte_in_word = byte_idx % 4u;
|
|
78
|
+
let byte_val = (qs[word_idx] >> (byte_in_word * 8u)) & 0xFFu;
|
|
79
|
+
|
|
80
|
+
if (use_upper) {
|
|
81
|
+
return (byte_val >> 4u) & 0xFu;
|
|
82
|
+
}
|
|
83
|
+
return byte_val & 0xFu;
|
|
84
|
+
}
|
|
85
|
+
|
|
86
|
+
const MAX_TILE_M: u32 = 4u;
|
|
87
|
+
const MAX_THREADS_PER_COL: u32 = 256u;
|
|
88
|
+
const MAX_SUBGROUPS_PER_ROW: u32 = 64u;
|
|
89
|
+
|
|
90
|
+
override TILE_M: u32 = 4u;
|
|
91
|
+
override THREADS_PER_COL: u32 = 256u;
|
|
92
|
+
|
|
93
|
+
var<workgroup> batched_wg_sums: array<f16, MAX_TILE_M * MAX_SUBGROUPS_PER_ROW>;
|
|
94
|
+
|
|
95
|
+
@compute @workgroup_size(THREADS_PER_COL, TILE_M, 1)
|
|
96
|
+
fn main_batched_f16acc_f16a(
|
|
97
|
+
@builtin(local_invocation_id) lid: vec3<u32>,
|
|
98
|
+
@builtin(workgroup_id) wg_id: vec3<u32>,
|
|
99
|
+
@builtin(subgroup_invocation_id) sg_id: u32,
|
|
100
|
+
@builtin(subgroup_size) sg_size: u32
|
|
101
|
+
) {
|
|
102
|
+
if (TILE_M > MAX_TILE_M || THREADS_PER_COL > MAX_THREADS_PER_COL) {
|
|
103
|
+
return;
|
|
104
|
+
}
|
|
105
|
+
|
|
106
|
+
let local_id = lid.x;
|
|
107
|
+
let row = wg_id.y * TILE_M + lid.y;
|
|
108
|
+
let col = wg_id.x;
|
|
109
|
+
let is_valid = row < u.M && col < u.N;
|
|
110
|
+
|
|
111
|
+
var partial_sum: f16 = f16(0.0);
|
|
112
|
+
|
|
113
|
+
if (is_valid) {
|
|
114
|
+
let num_blocks = u.num_blocks_per_row;
|
|
115
|
+
let total_subblocks = num_blocks * 8u;
|
|
116
|
+
for (var partial: u32 = local_id; partial < total_subblocks; partial = partial + THREADS_PER_COL) {
|
|
117
|
+
let b = partial / 8u;
|
|
118
|
+
let sb = partial % 8u;
|
|
119
|
+
let block = B_q4k[col * num_blocks + b];
|
|
120
|
+
let d = unpack_f16_lo(block.d_dmin);
|
|
121
|
+
let dmin = unpack_f16_hi(block.d_dmin);
|
|
122
|
+
let k_base = b * QK_K;
|
|
123
|
+
|
|
124
|
+
let sm = get_scale_min_k4(block.scales, sb);
|
|
125
|
+
let scale = d * f16(sm.x);
|
|
126
|
+
let min_val = dmin * f16(sm.y);
|
|
127
|
+
|
|
128
|
+
for (var i: u32 = 0u; i < SUBBLOCK_SIZE; i = i + 1u) {
|
|
129
|
+
let elem = sb * SUBBLOCK_SIZE + i;
|
|
130
|
+
let k = k_base + elem;
|
|
131
|
+
if (k < u.K) {
|
|
132
|
+
let a_val = A[row * u.K + k];
|
|
133
|
+
let q = get_q4(block.qs, elem);
|
|
134
|
+
let w = scale * f16(q) - min_val;
|
|
135
|
+
partial_sum = partial_sum + a_val * w;
|
|
136
|
+
}
|
|
137
|
+
}
|
|
138
|
+
}
|
|
139
|
+
}
|
|
140
|
+
|
|
141
|
+
let sg_sum = subgroupAdd(partial_sum);
|
|
142
|
+
let num_subgroups = (THREADS_PER_COL + sg_size - 1u) / sg_size;
|
|
143
|
+
|
|
144
|
+
if (sg_id == 0u && local_id < THREADS_PER_COL) {
|
|
145
|
+
let sg_idx = local_id / sg_size;
|
|
146
|
+
batched_wg_sums[lid.y * MAX_SUBGROUPS_PER_ROW + sg_idx] = sg_sum;
|
|
147
|
+
}
|
|
148
|
+
|
|
149
|
+
workgroupBarrier();
|
|
150
|
+
|
|
151
|
+
if (local_id == 0u && is_valid) {
|
|
152
|
+
var final_sum: f16 = f16(0.0);
|
|
153
|
+
for (var i: u32 = 0u; i < num_subgroups; i = i + 1u) {
|
|
154
|
+
final_sum = final_sum + batched_wg_sums[lid.y * MAX_SUBGROUPS_PER_ROW + i];
|
|
155
|
+
}
|
|
156
|
+
C_f16[row * u.N + col] = final_sum * f16(u.alpha);
|
|
157
|
+
}
|
|
158
|
+
}
|
|
@@ -6,7 +6,6 @@ enable f16;
|
|
|
6
6
|
enable subgroups;
|
|
7
7
|
|
|
8
8
|
const QK_K: u32 = 256u;
|
|
9
|
-
const BLOCK_SIZE: u32 = 144u;
|
|
10
9
|
const SUBBLOCK_SIZE: u32 = 32u;
|
|
11
10
|
|
|
12
11
|
override WORKGROUP_SIZE: u32 = 256u;
|
|
@@ -34,12 +33,12 @@ struct Q4KBlock {
|
|
|
34
33
|
@group(0) @binding(2) var<storage, read> B_q4k: array<Q4KBlock>;
|
|
35
34
|
@group(0) @binding(4) var<storage, read_write> C_f16: array<f16>;
|
|
36
35
|
|
|
37
|
-
fn unpack_f16_lo(packed: u32) ->
|
|
38
|
-
return unpack2x16float(packed).x;
|
|
36
|
+
fn unpack_f16_lo(packed: u32) -> f16 {
|
|
37
|
+
return f16(unpack2x16float(packed).x);
|
|
39
38
|
}
|
|
40
39
|
|
|
41
|
-
fn unpack_f16_hi(packed: u32) ->
|
|
42
|
-
return unpack2x16float(packed).y;
|
|
40
|
+
fn unpack_f16_hi(packed: u32) -> f16 {
|
|
41
|
+
return f16(unpack2x16float(packed).y);
|
|
43
42
|
}
|
|
44
43
|
|
|
45
44
|
fn get_scale_byte(scales: array<u32, 3>, byte_idx: u32) -> u32 {
|
|
@@ -86,7 +85,7 @@ fn get_q4(qs: array<u32, 32>, idx: u32) -> u32 {
|
|
|
86
85
|
override COLS_PER_WG: u32 = 32u;
|
|
87
86
|
override THREADS_PER_COL_GEMV: u32 = 8u;
|
|
88
87
|
|
|
89
|
-
var<workgroup> multicol_sums: array<
|
|
88
|
+
var<workgroup> multicol_sums: array<f16, MAX_WORKGROUP_SIZE>;
|
|
90
89
|
|
|
91
90
|
@compute @workgroup_size(WORKGROUP_SIZE, 1, 1)
|
|
92
91
|
fn main_multicol_f16a(
|
|
@@ -101,7 +100,7 @@ fn main_multicol_f16a(
|
|
|
101
100
|
let col = wg_id.x * COLS_PER_WG + col_in_wg;
|
|
102
101
|
let is_valid = col < u.N;
|
|
103
102
|
|
|
104
|
-
var partial_sum:
|
|
103
|
+
var partial_sum: f16 = f16(0.0);
|
|
105
104
|
|
|
106
105
|
if (is_valid) {
|
|
107
106
|
let num_blocks = u.num_blocks_per_row;
|
|
@@ -115,26 +114,26 @@ fn main_multicol_f16a(
|
|
|
115
114
|
|
|
116
115
|
for (var sb: u32 = 0u; sb < 8u; sb = sb + 1u) {
|
|
117
116
|
let sm = get_scale_min_k4(block.scales, sb);
|
|
118
|
-
let scale = d *
|
|
119
|
-
let min_val = dmin *
|
|
117
|
+
let scale = d * f16(sm.x);
|
|
118
|
+
let min_val = dmin * f16(sm.y);
|
|
120
119
|
let sb_base = sb * SUBBLOCK_SIZE;
|
|
121
120
|
|
|
122
121
|
for (var i: u32 = 0u; i < SUBBLOCK_SIZE; i = i + 4u) {
|
|
123
122
|
let k0 = k_base + sb_base + i;
|
|
124
|
-
let a0 =
|
|
125
|
-
let a1 =
|
|
126
|
-
let a2 =
|
|
127
|
-
let a3 =
|
|
123
|
+
let a0 = A[k0];
|
|
124
|
+
let a1 = A[k0 + 1u];
|
|
125
|
+
let a2 = A[k0 + 2u];
|
|
126
|
+
let a3 = A[k0 + 3u];
|
|
128
127
|
|
|
129
128
|
let q0 = get_q4(block.qs, sb_base + i);
|
|
130
129
|
let q1 = get_q4(block.qs, sb_base + i + 1u);
|
|
131
130
|
let q2 = get_q4(block.qs, sb_base + i + 2u);
|
|
132
131
|
let q3 = get_q4(block.qs, sb_base + i + 3u);
|
|
133
132
|
|
|
134
|
-
let w0 = scale *
|
|
135
|
-
let w1 = scale *
|
|
136
|
-
let w2 = scale *
|
|
137
|
-
let w3 = scale *
|
|
133
|
+
let w0 = scale * f16(q0) - min_val;
|
|
134
|
+
let w1 = scale * f16(q1) - min_val;
|
|
135
|
+
let w2 = scale * f16(q2) - min_val;
|
|
136
|
+
let w3 = scale * f16(q3) - min_val;
|
|
138
137
|
|
|
139
138
|
partial_sum = partial_sum + a0 * w0 + a1 * w1 + a2 * w2 + a3 * w3;
|
|
140
139
|
}
|
|
@@ -155,8 +154,8 @@ fn main_multicol_f16a(
|
|
|
155
154
|
break;
|
|
156
155
|
}
|
|
157
156
|
let sm = get_scale_min_k4(block.scales, sb);
|
|
158
|
-
let scale = d *
|
|
159
|
-
let min_val = dmin *
|
|
157
|
+
let scale = d * f16(sm.x);
|
|
158
|
+
let min_val = dmin * f16(sm.y);
|
|
160
159
|
|
|
161
160
|
for (var i: u32 = 0u; i < SUBBLOCK_SIZE; i = i + 4u) {
|
|
162
161
|
let k0 = k_base + sb_base + i;
|
|
@@ -164,24 +163,24 @@ fn main_multicol_f16a(
|
|
|
164
163
|
let k2 = k0 + 2u;
|
|
165
164
|
let k3 = k0 + 3u;
|
|
166
165
|
|
|
167
|
-
var a0:
|
|
168
|
-
var a1:
|
|
169
|
-
var a2:
|
|
170
|
-
var a3:
|
|
171
|
-
if (k0 < u.K) { a0 =
|
|
172
|
-
if (k1 < u.K) { a1 =
|
|
173
|
-
if (k2 < u.K) { a2 =
|
|
174
|
-
if (k3 < u.K) { a3 =
|
|
166
|
+
var a0: f16 = f16(0.0);
|
|
167
|
+
var a1: f16 = f16(0.0);
|
|
168
|
+
var a2: f16 = f16(0.0);
|
|
169
|
+
var a3: f16 = f16(0.0);
|
|
170
|
+
if (k0 < u.K) { a0 = A[k0]; }
|
|
171
|
+
if (k1 < u.K) { a1 = A[k1]; }
|
|
172
|
+
if (k2 < u.K) { a2 = A[k2]; }
|
|
173
|
+
if (k3 < u.K) { a3 = A[k3]; }
|
|
175
174
|
|
|
176
175
|
let q0 = get_q4(block.qs, sb_base + i);
|
|
177
176
|
let q1 = get_q4(block.qs, sb_base + i + 1u);
|
|
178
177
|
let q2 = get_q4(block.qs, sb_base + i + 2u);
|
|
179
178
|
let q3 = get_q4(block.qs, sb_base + i + 3u);
|
|
180
179
|
|
|
181
|
-
let w0 = scale *
|
|
182
|
-
let w1 = scale *
|
|
183
|
-
let w2 = scale *
|
|
184
|
-
let w3 = scale *
|
|
180
|
+
let w0 = scale * f16(q0) - min_val;
|
|
181
|
+
let w1 = scale * f16(q1) - min_val;
|
|
182
|
+
let w2 = scale * f16(q2) - min_val;
|
|
183
|
+
let w3 = scale * f16(q3) - min_val;
|
|
185
184
|
|
|
186
185
|
partial_sum = partial_sum + a0 * w0 + a1 * w1 + a2 * w2 + a3 * w3;
|
|
187
186
|
}
|
|
@@ -194,11 +193,11 @@ fn main_multicol_f16a(
|
|
|
194
193
|
workgroupBarrier();
|
|
195
194
|
|
|
196
195
|
if (tid_in_col == 0u && is_valid) {
|
|
197
|
-
var final_sum:
|
|
196
|
+
var final_sum: f16 = f16(0.0);
|
|
198
197
|
let base = col_in_wg * THREADS_PER_COL_GEMV;
|
|
199
198
|
for (var i: u32 = 0u; i < THREADS_PER_COL_GEMV; i = i + 1u) {
|
|
200
199
|
final_sum = final_sum + multicol_sums[base + i];
|
|
201
200
|
}
|
|
202
|
-
C_f16[col] =
|
|
201
|
+
C_f16[col] = final_sum * f16(u.alpha);
|
|
203
202
|
}
|
|
204
203
|
}
|
|
@@ -0,0 +1,209 @@
|
|
|
1
|
+
// Fused Q4_K Matmul Kernel - W4A16 (F16 IO, F32 accumulator, multicol)
|
|
2
|
+
//
|
|
3
|
+
// Computes C_f16 = A * dequant(B_q4k) for M=1 GEMV with f16 activations.
|
|
4
|
+
// Identical to fused_matmul_q4_multicol_f16a.wgsl except the per-thread
|
|
5
|
+
// accumulator and shared-memory reduction run in f32 instead of f16. On
|
|
6
|
+
// AMD RDNA3 the f16-accumulator path of the f16a kernel measured ~60%
|
|
7
|
+
// slower than this f32-accumulator twin at FFN gemv shape. Q4K weight
|
|
8
|
+
// reads stay 4-bit (bandwidth win) and IO stays f16 (memory layout
|
|
9
|
+
// unchanged); only the reduction precision changes.
|
|
10
|
+
|
|
11
|
+
enable f16;
|
|
12
|
+
enable subgroups;
|
|
13
|
+
|
|
14
|
+
const QK_K: u32 = 256u;
|
|
15
|
+
const SUBBLOCK_SIZE: u32 = 32u;
|
|
16
|
+
|
|
17
|
+
override WORKGROUP_SIZE: u32 = 256u;
|
|
18
|
+
const MAX_WORKGROUP_SIZE: u32 = 256u;
|
|
19
|
+
|
|
20
|
+
struct Uniforms {
|
|
21
|
+
M: u32,
|
|
22
|
+
N: u32,
|
|
23
|
+
K: u32,
|
|
24
|
+
alpha: f32,
|
|
25
|
+
num_blocks_per_row: u32,
|
|
26
|
+
_pad0: u32,
|
|
27
|
+
_pad1: u32,
|
|
28
|
+
_pad2: u32,
|
|
29
|
+
}
|
|
30
|
+
|
|
31
|
+
struct Q4KBlock {
|
|
32
|
+
d_dmin: u32,
|
|
33
|
+
scales: array<u32, 3>,
|
|
34
|
+
qs: array<u32, 32>,
|
|
35
|
+
}
|
|
36
|
+
|
|
37
|
+
@group(0) @binding(0) var<uniform> u: Uniforms;
|
|
38
|
+
@group(0) @binding(1) var<storage, read> A: array<f16>;
|
|
39
|
+
@group(0) @binding(2) var<storage, read> B_q4k: array<Q4KBlock>;
|
|
40
|
+
@group(0) @binding(4) var<storage, read_write> C_f16: array<f16>;
|
|
41
|
+
|
|
42
|
+
fn unpack_f16_lo(packed: u32) -> f16 {
|
|
43
|
+
return f16(unpack2x16float(packed).x);
|
|
44
|
+
}
|
|
45
|
+
|
|
46
|
+
fn unpack_f16_hi(packed: u32) -> f16 {
|
|
47
|
+
return f16(unpack2x16float(packed).y);
|
|
48
|
+
}
|
|
49
|
+
|
|
50
|
+
fn get_scale_byte(scales: array<u32, 3>, byte_idx: u32) -> u32 {
|
|
51
|
+
let word_idx = byte_idx / 4u;
|
|
52
|
+
let byte_in_word = byte_idx % 4u;
|
|
53
|
+
return (scales[word_idx] >> (byte_in_word * 8u)) & 0xFFu;
|
|
54
|
+
}
|
|
55
|
+
|
|
56
|
+
fn get_scale_min_k4(scales: array<u32, 3>, j: u32) -> vec2<u32> {
|
|
57
|
+
var sc: u32;
|
|
58
|
+
var mn: u32;
|
|
59
|
+
|
|
60
|
+
if (j < 4u) {
|
|
61
|
+
sc = get_scale_byte(scales, j) & 63u;
|
|
62
|
+
mn = get_scale_byte(scales, j + 4u) & 63u;
|
|
63
|
+
} else {
|
|
64
|
+
let q_j = get_scale_byte(scales, j + 4u);
|
|
65
|
+
let q_lo = get_scale_byte(scales, j - 4u);
|
|
66
|
+
let q_hi = get_scale_byte(scales, j);
|
|
67
|
+
sc = (q_j & 0xFu) | ((q_lo >> 6u) << 4u);
|
|
68
|
+
mn = (q_j >> 4u) | ((q_hi >> 6u) << 4u);
|
|
69
|
+
}
|
|
70
|
+
return vec2<u32>(sc, mn);
|
|
71
|
+
}
|
|
72
|
+
|
|
73
|
+
fn get_q4(qs: array<u32, 32>, idx: u32) -> u32 {
|
|
74
|
+
let chunk = idx / 64u;
|
|
75
|
+
let pos_in_chunk = idx % 64u;
|
|
76
|
+
let use_upper = pos_in_chunk >= 32u;
|
|
77
|
+
let byte_in_range = select(pos_in_chunk, pos_in_chunk - 32u, use_upper);
|
|
78
|
+
let byte_idx = chunk * 32u + byte_in_range;
|
|
79
|
+
|
|
80
|
+
let word_idx = byte_idx / 4u;
|
|
81
|
+
let byte_in_word = byte_idx % 4u;
|
|
82
|
+
let byte_val = (qs[word_idx] >> (byte_in_word * 8u)) & 0xFFu;
|
|
83
|
+
|
|
84
|
+
if (use_upper) {
|
|
85
|
+
return (byte_val >> 4u) & 0xFu;
|
|
86
|
+
} else {
|
|
87
|
+
return byte_val & 0xFu;
|
|
88
|
+
}
|
|
89
|
+
}
|
|
90
|
+
|
|
91
|
+
override COLS_PER_WG: u32 = 32u;
|
|
92
|
+
override THREADS_PER_COL_GEMV: u32 = 8u;
|
|
93
|
+
|
|
94
|
+
var<workgroup> multicol_sums: array<f32, MAX_WORKGROUP_SIZE>;
|
|
95
|
+
|
|
96
|
+
@compute @workgroup_size(WORKGROUP_SIZE, 1, 1)
|
|
97
|
+
fn main_multicol_f16a_f32acc(
|
|
98
|
+
@builtin(local_invocation_id) lid: vec3<u32>,
|
|
99
|
+
@builtin(workgroup_id) wg_id: vec3<u32>,
|
|
100
|
+
@builtin(subgroup_invocation_id) sg_id: u32,
|
|
101
|
+
@builtin(subgroup_size) sg_size: u32
|
|
102
|
+
) {
|
|
103
|
+
let local_id = lid.x;
|
|
104
|
+
let col_in_wg = local_id / THREADS_PER_COL_GEMV;
|
|
105
|
+
let tid_in_col = local_id % THREADS_PER_COL_GEMV;
|
|
106
|
+
let col = wg_id.x * COLS_PER_WG + col_in_wg;
|
|
107
|
+
let is_valid = col < u.N;
|
|
108
|
+
|
|
109
|
+
var partial_sum: f32 = 0.0;
|
|
110
|
+
|
|
111
|
+
if (is_valid) {
|
|
112
|
+
let num_blocks = u.num_blocks_per_row;
|
|
113
|
+
let tail_size = u.K & 255u;
|
|
114
|
+
let full_blocks = num_blocks - select(0u, 1u, tail_size > 0u);
|
|
115
|
+
for (var b: u32 = tid_in_col; b < full_blocks; b = b + THREADS_PER_COL_GEMV) {
|
|
116
|
+
let block = B_q4k[col * num_blocks + b];
|
|
117
|
+
let d = unpack_f16_lo(block.d_dmin);
|
|
118
|
+
let dmin = unpack_f16_hi(block.d_dmin);
|
|
119
|
+
let k_base = b * QK_K;
|
|
120
|
+
|
|
121
|
+
for (var sb: u32 = 0u; sb < 8u; sb = sb + 1u) {
|
|
122
|
+
let sm = get_scale_min_k4(block.scales, sb);
|
|
123
|
+
let scale = d * f16(sm.x);
|
|
124
|
+
let min_val = dmin * f16(sm.y);
|
|
125
|
+
let sb_base = sb * SUBBLOCK_SIZE;
|
|
126
|
+
|
|
127
|
+
for (var i: u32 = 0u; i < SUBBLOCK_SIZE; i = i + 4u) {
|
|
128
|
+
let k0 = k_base + sb_base + i;
|
|
129
|
+
let a0 = A[k0];
|
|
130
|
+
let a1 = A[k0 + 1u];
|
|
131
|
+
let a2 = A[k0 + 2u];
|
|
132
|
+
let a3 = A[k0 + 3u];
|
|
133
|
+
|
|
134
|
+
let q0 = get_q4(block.qs, sb_base + i);
|
|
135
|
+
let q1 = get_q4(block.qs, sb_base + i + 1u);
|
|
136
|
+
let q2 = get_q4(block.qs, sb_base + i + 2u);
|
|
137
|
+
let q3 = get_q4(block.qs, sb_base + i + 3u);
|
|
138
|
+
|
|
139
|
+
let w0 = scale * f16(q0) - min_val;
|
|
140
|
+
let w1 = scale * f16(q1) - min_val;
|
|
141
|
+
let w2 = scale * f16(q2) - min_val;
|
|
142
|
+
let w3 = scale * f16(q3) - min_val;
|
|
143
|
+
|
|
144
|
+
partial_sum = partial_sum + f32(a0) * f32(w0) + f32(a1) * f32(w1) + f32(a2) * f32(w2) + f32(a3) * f32(w3);
|
|
145
|
+
}
|
|
146
|
+
}
|
|
147
|
+
}
|
|
148
|
+
|
|
149
|
+
if (tail_size > 0u) {
|
|
150
|
+
let tail_block = full_blocks;
|
|
151
|
+
if (tail_block % THREADS_PER_COL_GEMV == tid_in_col) {
|
|
152
|
+
let block = B_q4k[col * num_blocks + tail_block];
|
|
153
|
+
let d = unpack_f16_lo(block.d_dmin);
|
|
154
|
+
let dmin = unpack_f16_hi(block.d_dmin);
|
|
155
|
+
let k_base = tail_block * QK_K;
|
|
156
|
+
|
|
157
|
+
for (var sb: u32 = 0u; sb < 8u; sb = sb + 1u) {
|
|
158
|
+
let sb_base = sb * SUBBLOCK_SIZE;
|
|
159
|
+
if (sb_base >= tail_size) {
|
|
160
|
+
break;
|
|
161
|
+
}
|
|
162
|
+
let sm = get_scale_min_k4(block.scales, sb);
|
|
163
|
+
let scale = d * f16(sm.x);
|
|
164
|
+
let min_val = dmin * f16(sm.y);
|
|
165
|
+
|
|
166
|
+
for (var i: u32 = 0u; i < SUBBLOCK_SIZE; i = i + 4u) {
|
|
167
|
+
let k0 = k_base + sb_base + i;
|
|
168
|
+
let k1 = k0 + 1u;
|
|
169
|
+
let k2 = k0 + 2u;
|
|
170
|
+
let k3 = k0 + 3u;
|
|
171
|
+
|
|
172
|
+
var a0: f16 = f16(0.0);
|
|
173
|
+
var a1: f16 = f16(0.0);
|
|
174
|
+
var a2: f16 = f16(0.0);
|
|
175
|
+
var a3: f16 = f16(0.0);
|
|
176
|
+
if (k0 < u.K) { a0 = A[k0]; }
|
|
177
|
+
if (k1 < u.K) { a1 = A[k1]; }
|
|
178
|
+
if (k2 < u.K) { a2 = A[k2]; }
|
|
179
|
+
if (k3 < u.K) { a3 = A[k3]; }
|
|
180
|
+
|
|
181
|
+
let q0 = get_q4(block.qs, sb_base + i);
|
|
182
|
+
let q1 = get_q4(block.qs, sb_base + i + 1u);
|
|
183
|
+
let q2 = get_q4(block.qs, sb_base + i + 2u);
|
|
184
|
+
let q3 = get_q4(block.qs, sb_base + i + 3u);
|
|
185
|
+
|
|
186
|
+
let w0 = scale * f16(q0) - min_val;
|
|
187
|
+
let w1 = scale * f16(q1) - min_val;
|
|
188
|
+
let w2 = scale * f16(q2) - min_val;
|
|
189
|
+
let w3 = scale * f16(q3) - min_val;
|
|
190
|
+
|
|
191
|
+
partial_sum = partial_sum + f32(a0) * f32(w0) + f32(a1) * f32(w1) + f32(a2) * f32(w2) + f32(a3) * f32(w3);
|
|
192
|
+
}
|
|
193
|
+
}
|
|
194
|
+
}
|
|
195
|
+
}
|
|
196
|
+
}
|
|
197
|
+
|
|
198
|
+
multicol_sums[local_id] = partial_sum;
|
|
199
|
+
workgroupBarrier();
|
|
200
|
+
|
|
201
|
+
if (tid_in_col == 0u && is_valid) {
|
|
202
|
+
var final_sum: f32 = 0.0;
|
|
203
|
+
let base = col_in_wg * THREADS_PER_COL_GEMV;
|
|
204
|
+
for (var i: u32 = 0u; i < THREADS_PER_COL_GEMV; i = i + 1u) {
|
|
205
|
+
final_sum = final_sum + multicol_sums[base + i];
|
|
206
|
+
}
|
|
207
|
+
C_f16[col] = f16(final_sum * u.alpha);
|
|
208
|
+
}
|
|
209
|
+
}
|