doppler-gpu 0.4.2 → 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 +19 -15
- package/package.json +77 -29
- 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 +4 -2
- 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 +64 -6
- 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 +367 -61
- 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 +524 -350
- 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 +1 -1
- 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/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 +12 -2
- package/src/config/schema/doppler.schema.js +108 -8
- 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 +52 -1
- package/src/config/schema/index.js +14 -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 +68 -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 +76 -0
- 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/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 +107 -0
- package/src/formats/rdrr/validation.js +245 -0
- package/src/formats/tflite/types.d.ts +0 -31
- package/src/formats/tflite/types.js +5 -5
- package/src/gpu/command-recorder.d.ts +3 -21
- 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.wgsl +3 -1
- 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 +6 -5
- package/src/gpu/kernels/index.js +3 -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 +12 -0
- package/src/inference/browser-harness-text-helpers.js +223 -14
- package/src/inference/browser-harness.d.ts +0 -29
- package/src/inference/browser-harness.js +264 -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 +17 -0
- 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/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 +22 -27
- package/src/inference/pipelines/text/embed.js +0 -39
- 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 +0 -2
- package/src/inference/pipelines/text/generator-helpers.js +4 -5
- 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 +8 -3
- package/src/inference/pipelines/text/generator.d.ts +0 -6
- package/src/inference/pipelines/text/generator.js +223 -547
- 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 +5 -2
- package/src/inference/pipelines/text/init.js +183 -724
- 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.js +2 -0
- 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 +31 -6
- package/src/inference/pipelines/text/probes.d.ts +3 -6
- package/src/inference/pipelines/text/probes.js +120 -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 +3 -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 +223 -27
- 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 +13 -3
- package/src/loader/loader-types.d.ts +2 -0
- package/src/loader/manifest-config.js +25 -1
- package/src/loader/memory-monitor.d.ts +17 -5
- 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.js +6 -1
- package/src/loader/tensors/tensor-loader.d.ts +2 -9
- package/src/loader/tensors/tensor-loader.js +84 -1
- 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 +11 -3
- package/src/storage/artifact-storage-context.js +231 -26
- 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 +11 -0
- package/src/storage/quickstart-downloader.js +44 -67
- 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 +21 -24
- 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 +1 -10
- 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 +2 -0
- package/src/tooling/source-runtime-bundle.js +3 -22
- 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 +1 -7
- 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
|
@@ -1,738 +0,0 @@
|
|
|
1
|
-
import { getDevice } from '../../../gpu/device.js';
|
|
2
|
-
import { createTensor } from '../../../gpu/tensor.js';
|
|
3
|
-
import { getBuffer } from '../../../gpu/weight-buffer.js';
|
|
4
|
-
import { acquireBuffer } from '../../../memory/buffer-pool.js';
|
|
5
|
-
import {
|
|
6
|
-
runConv2D,
|
|
7
|
-
runDepthwiseConv2D,
|
|
8
|
-
runGroupedPointwiseConv2D,
|
|
9
|
-
runLayerNorm,
|
|
10
|
-
runRMSNorm,
|
|
11
|
-
runMatmul,
|
|
12
|
-
runAttention,
|
|
13
|
-
runSiLU,
|
|
14
|
-
runSiLURowSplit,
|
|
15
|
-
runResidualAdd,
|
|
16
|
-
runBiasAdd,
|
|
17
|
-
runModulate,
|
|
18
|
-
runSanaLinearAttention,
|
|
19
|
-
recordConv2D,
|
|
20
|
-
recordDepthwiseConv2D,
|
|
21
|
-
recordGroupedPointwiseConv2D,
|
|
22
|
-
recordLayerNorm,
|
|
23
|
-
recordRMSNorm,
|
|
24
|
-
recordMatmul,
|
|
25
|
-
recordAttention,
|
|
26
|
-
recordSiLU,
|
|
27
|
-
recordSiLURowSplit,
|
|
28
|
-
recordResidualAdd,
|
|
29
|
-
recordBiasAdd,
|
|
30
|
-
recordModulate,
|
|
31
|
-
recordSanaLinearAttention,
|
|
32
|
-
} from '../../../gpu/kernels/index.js';
|
|
33
|
-
import { log } from '../../../debug/index.js';
|
|
34
|
-
import {
|
|
35
|
-
resolveDiffusionActivationDtype,
|
|
36
|
-
createDiffusionBufferReleaser,
|
|
37
|
-
createDiffusionBufferDestroyer,
|
|
38
|
-
normalizeDiffusionMatmulLocationDtype,
|
|
39
|
-
inferDiffusionMatmulDtypeFromBuffer,
|
|
40
|
-
expectDiffusionWeight,
|
|
41
|
-
} from './helpers.js';
|
|
42
|
-
|
|
43
|
-
function reshapeTensor(tensor, shape, label) {
|
|
44
|
-
return createTensor(tensor.buffer, tensor.dtype, shape, label ?? tensor.label);
|
|
45
|
-
}
|
|
46
|
-
|
|
47
|
-
function getWeight(weightsEntry, name) {
|
|
48
|
-
return weightsEntry?.weights?.get(`transformer.${name}`) ?? null;
|
|
49
|
-
}
|
|
50
|
-
|
|
51
|
-
function getWeightShape(weightsEntry, name) {
|
|
52
|
-
return weightsEntry?.shapes?.get(`transformer.${name}`) ?? null;
|
|
53
|
-
}
|
|
54
|
-
|
|
55
|
-
function getWeightDtype(weightsEntry, name) {
|
|
56
|
-
return weightsEntry?.dtypes?.get(`transformer.${name}`) ?? null;
|
|
57
|
-
}
|
|
58
|
-
|
|
59
|
-
function createKernelOps(recorder) {
|
|
60
|
-
if (!recorder) {
|
|
61
|
-
return {
|
|
62
|
-
conv2d: runConv2D,
|
|
63
|
-
depthwiseConv2d: runDepthwiseConv2D,
|
|
64
|
-
groupedPointwiseConv2d: runGroupedPointwiseConv2D,
|
|
65
|
-
layerNorm: runLayerNorm,
|
|
66
|
-
rmsNorm: runRMSNorm,
|
|
67
|
-
attention: runAttention,
|
|
68
|
-
silu: runSiLU,
|
|
69
|
-
siluRowSplit: runSiLURowSplit,
|
|
70
|
-
residualAdd: runResidualAdd,
|
|
71
|
-
biasAdd: runBiasAdd,
|
|
72
|
-
modulate: runModulate,
|
|
73
|
-
sanaLinearAttention: runSanaLinearAttention,
|
|
74
|
-
};
|
|
75
|
-
}
|
|
76
|
-
return {
|
|
77
|
-
conv2d: (...args) => recordConv2D(recorder, ...args),
|
|
78
|
-
depthwiseConv2d: (...args) => recordDepthwiseConv2D(recorder, ...args),
|
|
79
|
-
groupedPointwiseConv2d: (...args) => recordGroupedPointwiseConv2D(recorder, ...args),
|
|
80
|
-
layerNorm: (...args) => recordLayerNorm(recorder, ...args),
|
|
81
|
-
rmsNorm: (...args) => recordRMSNorm(recorder, ...args),
|
|
82
|
-
attention: (...args) => recordAttention(recorder, ...args),
|
|
83
|
-
silu: (...args) => recordSiLU(recorder, ...args),
|
|
84
|
-
siluRowSplit: (...args) => recordSiLURowSplit(recorder, ...args),
|
|
85
|
-
residualAdd: (...args) => recordResidualAdd(recorder, ...args),
|
|
86
|
-
biasAdd: (...args) => recordBiasAdd(recorder, ...args),
|
|
87
|
-
modulate: (...args) => recordModulate(recorder, ...args),
|
|
88
|
-
sanaLinearAttention: (...args) => recordSanaLinearAttention(recorder, ...args),
|
|
89
|
-
};
|
|
90
|
-
}
|
|
91
|
-
|
|
92
|
-
function createVectorBuffer(device, values, label) {
|
|
93
|
-
const buffer = acquireBuffer(values.byteLength, undefined, label);
|
|
94
|
-
device.queue.writeBuffer(buffer, 0, values);
|
|
95
|
-
return buffer;
|
|
96
|
-
}
|
|
97
|
-
|
|
98
|
-
function createBiasTensor(weight, size, label) {
|
|
99
|
-
if (!weight) return null;
|
|
100
|
-
return createTensor(getBuffer(weight), 'f32', [size], label);
|
|
101
|
-
}
|
|
102
|
-
|
|
103
|
-
function countMask(mask) {
|
|
104
|
-
if (!mask) return null;
|
|
105
|
-
let count = 0;
|
|
106
|
-
for (const value of mask) {
|
|
107
|
-
if (value) count += 1;
|
|
108
|
-
}
|
|
109
|
-
return count;
|
|
110
|
-
}
|
|
111
|
-
|
|
112
|
-
function trimContext(context, attentionMask) {
|
|
113
|
-
const validTokens = countMask(attentionMask);
|
|
114
|
-
if (!Number.isFinite(validTokens) || validTokens <= 0 || validTokens >= context.shape[0]) {
|
|
115
|
-
return context;
|
|
116
|
-
}
|
|
117
|
-
return createTensor(context.buffer, context.dtype, [validTokens, context.shape[1]], 'sana_trimmed_context');
|
|
118
|
-
}
|
|
119
|
-
|
|
120
|
-
function buildSinusoidalEmbedding(value, dim = 256) {
|
|
121
|
-
const half = Math.floor(dim / 2);
|
|
122
|
-
const out = new Float32Array(dim);
|
|
123
|
-
const maxPeriod = 10000;
|
|
124
|
-
for (let i = 0; i < half; i++) {
|
|
125
|
-
const freq = Math.exp(-Math.log(maxPeriod) * i / half);
|
|
126
|
-
const angle = value * freq;
|
|
127
|
-
out[2 * i] = Math.cos(angle);
|
|
128
|
-
out[2 * i + 1] = Math.sin(angle);
|
|
129
|
-
}
|
|
130
|
-
return out;
|
|
131
|
-
}
|
|
132
|
-
|
|
133
|
-
function resolveMatmulDtype(weightsEntry, name, N, K) {
|
|
134
|
-
const weight = getWeight(weightsEntry, name);
|
|
135
|
-
const preferred = normalizeDiffusionMatmulLocationDtype(getWeightDtype(weightsEntry, name));
|
|
136
|
-
return inferDiffusionMatmulDtypeFromBuffer(weight, N, K, preferred);
|
|
137
|
-
}
|
|
138
|
-
|
|
139
|
-
async function runMatmulResolved(input, weightsEntry, name, M, N, K, recorder, options = {}) {
|
|
140
|
-
const weight = expectDiffusionWeight(getWeight(weightsEntry, name), name);
|
|
141
|
-
const bDtype = resolveMatmulDtype(weightsEntry, name, N, K);
|
|
142
|
-
if (recorder) {
|
|
143
|
-
return recordMatmul(recorder, input, weight, M, N, K, { ...options, bDtype, transposeB: 'auto' });
|
|
144
|
-
}
|
|
145
|
-
return runMatmul(input, weight, M, N, K, { ...options, bDtype, transposeB: 'auto' });
|
|
146
|
-
}
|
|
147
|
-
|
|
148
|
-
async function runTwoLayerEmbedding(inputTensor, weightsEntry, prefix, outDim, recorder, runtime, ops, release) {
|
|
149
|
-
const activationDtype = resolveDiffusionActivationDtype(runtime);
|
|
150
|
-
|
|
151
|
-
let output = await runMatmulResolved(
|
|
152
|
-
inputTensor,
|
|
153
|
-
weightsEntry,
|
|
154
|
-
`${prefix}.linear_1.weight`,
|
|
155
|
-
1,
|
|
156
|
-
getWeightShape(weightsEntry, `${prefix}.linear_1.weight`)[0],
|
|
157
|
-
getWeightShape(weightsEntry, `${prefix}.linear_1.weight`)[1],
|
|
158
|
-
recorder,
|
|
159
|
-
{ outputDtype: activationDtype }
|
|
160
|
-
);
|
|
161
|
-
const bias1 = createBiasTensor(getWeight(weightsEntry, `${prefix}.linear_1.bias`), output.shape[1], `${prefix}_bias1`);
|
|
162
|
-
if (bias1) {
|
|
163
|
-
output = await ops.biasAdd(output, bias1, 1, output.shape[1]);
|
|
164
|
-
}
|
|
165
|
-
|
|
166
|
-
const act = await ops.silu(output, { size: output.shape[1], swigluLimit: null });
|
|
167
|
-
release(output.buffer);
|
|
168
|
-
|
|
169
|
-
let projected = await runMatmulResolved(
|
|
170
|
-
act,
|
|
171
|
-
weightsEntry,
|
|
172
|
-
`${prefix}.linear_2.weight`,
|
|
173
|
-
1,
|
|
174
|
-
outDim,
|
|
175
|
-
getWeightShape(weightsEntry, `${prefix}.linear_2.weight`)[1],
|
|
176
|
-
recorder,
|
|
177
|
-
{ outputDtype: activationDtype }
|
|
178
|
-
);
|
|
179
|
-
const bias2 = createBiasTensor(getWeight(weightsEntry, `${prefix}.linear_2.bias`), outDim, `${prefix}_bias2`);
|
|
180
|
-
if (bias2) {
|
|
181
|
-
projected = await ops.biasAdd(projected, bias2, 1, outDim);
|
|
182
|
-
}
|
|
183
|
-
release(act.buffer);
|
|
184
|
-
return projected;
|
|
185
|
-
}
|
|
186
|
-
|
|
187
|
-
export async function buildSanaTimestepConditioning(timestep, guidanceScale, weightsEntry, config, runtime, options = {}) {
|
|
188
|
-
const device = getDevice();
|
|
189
|
-
if (!device) {
|
|
190
|
-
throw new Error('Sana timestep conditioning requires a WebGPU device.');
|
|
191
|
-
}
|
|
192
|
-
const recorder = options.recorder ?? null;
|
|
193
|
-
const ops = createKernelOps(recorder);
|
|
194
|
-
const release = createDiffusionBufferReleaser(recorder);
|
|
195
|
-
const activationDtype = resolveDiffusionActivationDtype(runtime);
|
|
196
|
-
const hiddenSize = config.num_attention_heads * config.attention_head_dim;
|
|
197
|
-
|
|
198
|
-
const timeTensor = createTensor(
|
|
199
|
-
createVectorBuffer(device, buildSinusoidalEmbedding(timestep, 256), 'sana_timestep'),
|
|
200
|
-
activationDtype,
|
|
201
|
-
[1, 256],
|
|
202
|
-
'sana_timestep'
|
|
203
|
-
);
|
|
204
|
-
const timeEmbedding = await runTwoLayerEmbedding(
|
|
205
|
-
timeTensor,
|
|
206
|
-
weightsEntry,
|
|
207
|
-
'time_embed.timestep_embedder',
|
|
208
|
-
hiddenSize,
|
|
209
|
-
recorder,
|
|
210
|
-
runtime,
|
|
211
|
-
ops,
|
|
212
|
-
release
|
|
213
|
-
);
|
|
214
|
-
release(timeTensor.buffer);
|
|
215
|
-
|
|
216
|
-
let conditioning = timeEmbedding;
|
|
217
|
-
if (config.guidance_embeds === true) {
|
|
218
|
-
const guidanceTensor = createTensor(
|
|
219
|
-
createVectorBuffer(device, buildSinusoidalEmbedding(guidanceScale, 256), 'sana_guidance'),
|
|
220
|
-
activationDtype,
|
|
221
|
-
[1, 256],
|
|
222
|
-
'sana_guidance'
|
|
223
|
-
);
|
|
224
|
-
const guidanceEmbedding = await runTwoLayerEmbedding(
|
|
225
|
-
guidanceTensor,
|
|
226
|
-
weightsEntry,
|
|
227
|
-
'time_embed.guidance_embedder',
|
|
228
|
-
hiddenSize,
|
|
229
|
-
recorder,
|
|
230
|
-
runtime,
|
|
231
|
-
ops,
|
|
232
|
-
release
|
|
233
|
-
);
|
|
234
|
-
release(guidanceTensor.buffer);
|
|
235
|
-
conditioning = await ops.residualAdd(timeEmbedding, guidanceEmbedding, hiddenSize, { useVec4: true });
|
|
236
|
-
release(timeEmbedding.buffer);
|
|
237
|
-
release(guidanceEmbedding.buffer);
|
|
238
|
-
}
|
|
239
|
-
|
|
240
|
-
const conditioningAct = await ops.silu(conditioning, { size: hiddenSize, swigluLimit: null });
|
|
241
|
-
let modulation = await runMatmulResolved(
|
|
242
|
-
conditioningAct,
|
|
243
|
-
weightsEntry,
|
|
244
|
-
'time_embed.linear.weight',
|
|
245
|
-
1,
|
|
246
|
-
hiddenSize * 6,
|
|
247
|
-
hiddenSize,
|
|
248
|
-
recorder,
|
|
249
|
-
{ outputDtype: activationDtype }
|
|
250
|
-
);
|
|
251
|
-
const modulationBias = createBiasTensor(getWeight(weightsEntry, 'time_embed.linear.bias'), hiddenSize * 6, 'sana_time_linear_bias');
|
|
252
|
-
if (modulationBias) {
|
|
253
|
-
modulation = await ops.biasAdd(modulation, modulationBias, 1, hiddenSize * 6);
|
|
254
|
-
}
|
|
255
|
-
release(conditioningAct.buffer);
|
|
256
|
-
|
|
257
|
-
return {
|
|
258
|
-
modulation,
|
|
259
|
-
embeddedTimestep: conditioning,
|
|
260
|
-
};
|
|
261
|
-
}
|
|
262
|
-
|
|
263
|
-
export async function projectSanaContext(context, attentionMask, weightsEntry, config, runtime, options = {}) {
|
|
264
|
-
const recorder = options.recorder ?? null;
|
|
265
|
-
const ops = createKernelOps(recorder);
|
|
266
|
-
const release = createDiffusionBufferReleaser(recorder);
|
|
267
|
-
const trimmed = trimContext(context, attentionMask);
|
|
268
|
-
const tokenCount = trimmed.shape[0];
|
|
269
|
-
const inputDim = trimmed.shape[1];
|
|
270
|
-
const hiddenSize = config.num_attention_heads * config.attention_head_dim;
|
|
271
|
-
const activationDtype = resolveDiffusionActivationDtype(runtime);
|
|
272
|
-
|
|
273
|
-
let hidden = await runMatmulResolved(
|
|
274
|
-
trimmed,
|
|
275
|
-
weightsEntry,
|
|
276
|
-
'caption_projection.linear_1.weight',
|
|
277
|
-
tokenCount,
|
|
278
|
-
hiddenSize,
|
|
279
|
-
inputDim,
|
|
280
|
-
recorder,
|
|
281
|
-
{ outputDtype: activationDtype }
|
|
282
|
-
);
|
|
283
|
-
const bias1 = createBiasTensor(getWeight(weightsEntry, 'caption_projection.linear_1.bias'), hiddenSize, 'sana_caption_bias1');
|
|
284
|
-
if (bias1) {
|
|
285
|
-
hidden = await ops.biasAdd(hidden, bias1, tokenCount, hiddenSize);
|
|
286
|
-
}
|
|
287
|
-
|
|
288
|
-
// PixArtAlphaTextProjection uses GELU(tanh). Reuse the existing GeLU kernel here.
|
|
289
|
-
const { runGeLU, recordGeLU } = await import('../../../gpu/kernels/gelu.js');
|
|
290
|
-
const gelu = recorder
|
|
291
|
-
? (input, options = {}) => recordGeLU(recorder, input, options)
|
|
292
|
-
: runGeLU;
|
|
293
|
-
const activated = await gelu(hidden, { size: tokenCount * hiddenSize });
|
|
294
|
-
release(hidden.buffer);
|
|
295
|
-
|
|
296
|
-
let projected = await runMatmulResolved(
|
|
297
|
-
activated,
|
|
298
|
-
weightsEntry,
|
|
299
|
-
'caption_projection.linear_2.weight',
|
|
300
|
-
tokenCount,
|
|
301
|
-
hiddenSize,
|
|
302
|
-
hiddenSize,
|
|
303
|
-
recorder,
|
|
304
|
-
{ outputDtype: activationDtype }
|
|
305
|
-
);
|
|
306
|
-
const bias2 = createBiasTensor(getWeight(weightsEntry, 'caption_projection.linear_2.bias'), hiddenSize, 'sana_caption_bias2');
|
|
307
|
-
if (bias2) {
|
|
308
|
-
projected = await ops.biasAdd(projected, bias2, tokenCount, hiddenSize);
|
|
309
|
-
}
|
|
310
|
-
release(activated.buffer);
|
|
311
|
-
|
|
312
|
-
const normWeight = expectDiffusionWeight(getWeight(weightsEntry, 'caption_norm.weight'), 'caption_norm.weight');
|
|
313
|
-
const normed = await ops.rmsNorm(projected, getBuffer(normWeight), 1e-5, {
|
|
314
|
-
batchSize: tokenCount,
|
|
315
|
-
hiddenSize,
|
|
316
|
-
});
|
|
317
|
-
release(projected.buffer);
|
|
318
|
-
return normed;
|
|
319
|
-
}
|
|
320
|
-
|
|
321
|
-
async function duplicateVectorTensor(tensor, times, recorder) {
|
|
322
|
-
const device = getDevice();
|
|
323
|
-
const output = acquireBuffer(tensor.buffer.size * times, undefined, 'sana_duplicate_vector');
|
|
324
|
-
const encoder = recorder ? recorder.getEncoder() : device.createCommandEncoder();
|
|
325
|
-
for (let i = 0; i < times; i++) {
|
|
326
|
-
encoder.copyBufferToBuffer(tensor.buffer, 0, output, i * tensor.buffer.size, tensor.buffer.size);
|
|
327
|
-
}
|
|
328
|
-
if (!recorder) {
|
|
329
|
-
device.queue.submit([encoder.finish()]);
|
|
330
|
-
}
|
|
331
|
-
return createTensor(output, tensor.dtype, [1, tensor.shape[1] * times], 'sana_duplicate_vector');
|
|
332
|
-
}
|
|
333
|
-
|
|
334
|
-
async function buildLayerModulation(baseModulation, tableWeight, tableShape, recorder, ops, release) {
|
|
335
|
-
const segments = Array.isArray(tableShape) ? tableShape.reduce((acc, value) => acc * value, 1) : 0;
|
|
336
|
-
let combined = await ops.biasAdd(
|
|
337
|
-
baseModulation,
|
|
338
|
-
createTensor(getBuffer(tableWeight), baseModulation.dtype, [segments], 'sana_layer_table'),
|
|
339
|
-
1,
|
|
340
|
-
segments
|
|
341
|
-
);
|
|
342
|
-
return combined;
|
|
343
|
-
}
|
|
344
|
-
|
|
345
|
-
async function runSelfAttention(hiddenStates, layerIdx, weightsEntry, numTokens, hiddenSize, numHeads, headDim, eps, recorder, runtime, ops, release) {
|
|
346
|
-
let query = await runMatmulResolved(
|
|
347
|
-
hiddenStates,
|
|
348
|
-
weightsEntry,
|
|
349
|
-
`transformer_blocks.${layerIdx}.attn1.to_q.weight`,
|
|
350
|
-
numTokens,
|
|
351
|
-
hiddenSize,
|
|
352
|
-
hiddenSize,
|
|
353
|
-
recorder,
|
|
354
|
-
{ outputDtype: hiddenStates.dtype }
|
|
355
|
-
);
|
|
356
|
-
let key = await runMatmulResolved(
|
|
357
|
-
hiddenStates,
|
|
358
|
-
weightsEntry,
|
|
359
|
-
`transformer_blocks.${layerIdx}.attn1.to_k.weight`,
|
|
360
|
-
numTokens,
|
|
361
|
-
hiddenSize,
|
|
362
|
-
hiddenSize,
|
|
363
|
-
recorder,
|
|
364
|
-
{ outputDtype: hiddenStates.dtype }
|
|
365
|
-
);
|
|
366
|
-
let value = await runMatmulResolved(
|
|
367
|
-
hiddenStates,
|
|
368
|
-
weightsEntry,
|
|
369
|
-
`transformer_blocks.${layerIdx}.attn1.to_v.weight`,
|
|
370
|
-
numTokens,
|
|
371
|
-
hiddenSize,
|
|
372
|
-
hiddenSize,
|
|
373
|
-
recorder,
|
|
374
|
-
{ outputDtype: hiddenStates.dtype }
|
|
375
|
-
);
|
|
376
|
-
|
|
377
|
-
const normQ = getWeight(weightsEntry, `transformer_blocks.${layerIdx}.attn1.norm_q.weight`);
|
|
378
|
-
const normK = getWeight(weightsEntry, `transformer_blocks.${layerIdx}.attn1.norm_k.weight`);
|
|
379
|
-
if (normQ) {
|
|
380
|
-
const next = await ops.rmsNorm(query, getBuffer(normQ), eps, { batchSize: numTokens, hiddenSize });
|
|
381
|
-
release(query.buffer);
|
|
382
|
-
query = next;
|
|
383
|
-
}
|
|
384
|
-
if (normK) {
|
|
385
|
-
const next = await ops.rmsNorm(key, getBuffer(normK), eps, { batchSize: numTokens, hiddenSize });
|
|
386
|
-
release(key.buffer);
|
|
387
|
-
key = next;
|
|
388
|
-
}
|
|
389
|
-
|
|
390
|
-
let output = await ops.sanaLinearAttention(query, key, value, {
|
|
391
|
-
numHeads,
|
|
392
|
-
headDim,
|
|
393
|
-
numTokens,
|
|
394
|
-
hiddenSize,
|
|
395
|
-
eps: 1e-15,
|
|
396
|
-
});
|
|
397
|
-
release(query.buffer);
|
|
398
|
-
release(key.buffer);
|
|
399
|
-
release(value.buffer);
|
|
400
|
-
|
|
401
|
-
let projected = await runMatmulResolved(
|
|
402
|
-
output,
|
|
403
|
-
weightsEntry,
|
|
404
|
-
`transformer_blocks.${layerIdx}.attn1.to_out.0.weight`,
|
|
405
|
-
numTokens,
|
|
406
|
-
hiddenSize,
|
|
407
|
-
hiddenSize,
|
|
408
|
-
recorder,
|
|
409
|
-
{ outputDtype: hiddenStates.dtype }
|
|
410
|
-
);
|
|
411
|
-
release(output.buffer);
|
|
412
|
-
const outBias = createBiasTensor(getWeight(weightsEntry, `transformer_blocks.${layerIdx}.attn1.to_out.0.bias`), hiddenSize, 'sana_self_attn_bias');
|
|
413
|
-
if (outBias) {
|
|
414
|
-
projected = await ops.biasAdd(projected, outBias, numTokens, hiddenSize);
|
|
415
|
-
}
|
|
416
|
-
return projected;
|
|
417
|
-
}
|
|
418
|
-
|
|
419
|
-
async function runCrossAttention(hiddenStates, context, layerIdx, weightsEntry, numTokens, hiddenSize, config, recorder, ops, release) {
|
|
420
|
-
const qHeads = getWeightShape(weightsEntry, `transformer_blocks.${layerIdx}.attn2.to_q.weight`)[0];
|
|
421
|
-
const kHeads = getWeightShape(weightsEntry, `transformer_blocks.${layerIdx}.attn2.to_k.weight`)[0];
|
|
422
|
-
const vHeads = getWeightShape(weightsEntry, `transformer_blocks.${layerIdx}.attn2.to_v.weight`)[0];
|
|
423
|
-
const contextTokens = context.shape[0];
|
|
424
|
-
let query = await runMatmulResolved(
|
|
425
|
-
hiddenStates,
|
|
426
|
-
weightsEntry,
|
|
427
|
-
`transformer_blocks.${layerIdx}.attn2.to_q.weight`,
|
|
428
|
-
numTokens,
|
|
429
|
-
qHeads,
|
|
430
|
-
hiddenSize,
|
|
431
|
-
recorder,
|
|
432
|
-
{ outputDtype: hiddenStates.dtype }
|
|
433
|
-
);
|
|
434
|
-
let key = await runMatmulResolved(
|
|
435
|
-
context,
|
|
436
|
-
weightsEntry,
|
|
437
|
-
`transformer_blocks.${layerIdx}.attn2.to_k.weight`,
|
|
438
|
-
contextTokens,
|
|
439
|
-
kHeads,
|
|
440
|
-
context.shape[1],
|
|
441
|
-
recorder,
|
|
442
|
-
{ outputDtype: hiddenStates.dtype }
|
|
443
|
-
);
|
|
444
|
-
let value = await runMatmulResolved(
|
|
445
|
-
context,
|
|
446
|
-
weightsEntry,
|
|
447
|
-
`transformer_blocks.${layerIdx}.attn2.to_v.weight`,
|
|
448
|
-
contextTokens,
|
|
449
|
-
vHeads,
|
|
450
|
-
context.shape[1],
|
|
451
|
-
recorder,
|
|
452
|
-
{ outputDtype: hiddenStates.dtype }
|
|
453
|
-
);
|
|
454
|
-
|
|
455
|
-
const crossHeads = config.num_cross_attention_heads;
|
|
456
|
-
const headDim = config.cross_attention_head_dim;
|
|
457
|
-
const normQ = getWeight(weightsEntry, `transformer_blocks.${layerIdx}.attn2.norm_q.weight`);
|
|
458
|
-
const normK = getWeight(weightsEntry, `transformer_blocks.${layerIdx}.attn2.norm_k.weight`);
|
|
459
|
-
if (normQ) {
|
|
460
|
-
const next = await ops.rmsNorm(query, getBuffer(normQ), Number(config.norm_eps ?? 1e-6), {
|
|
461
|
-
batchSize: numTokens,
|
|
462
|
-
hiddenSize: qHeads,
|
|
463
|
-
});
|
|
464
|
-
release(query.buffer);
|
|
465
|
-
query = next;
|
|
466
|
-
}
|
|
467
|
-
if (normK) {
|
|
468
|
-
const next = await ops.rmsNorm(key, getBuffer(normK), Number(config.norm_eps ?? 1e-6), {
|
|
469
|
-
batchSize: contextTokens,
|
|
470
|
-
hiddenSize: kHeads,
|
|
471
|
-
});
|
|
472
|
-
release(key.buffer);
|
|
473
|
-
key = next;
|
|
474
|
-
}
|
|
475
|
-
|
|
476
|
-
const attention = await ops.attention(query, key, value, null, crossHeads, headDim, {
|
|
477
|
-
seqLen: numTokens,
|
|
478
|
-
kvLen: contextTokens,
|
|
479
|
-
numKVHeads: crossHeads,
|
|
480
|
-
causal: false,
|
|
481
|
-
});
|
|
482
|
-
release(query.buffer);
|
|
483
|
-
release(key.buffer);
|
|
484
|
-
release(value.buffer);
|
|
485
|
-
|
|
486
|
-
let projected = await runMatmulResolved(
|
|
487
|
-
attention,
|
|
488
|
-
weightsEntry,
|
|
489
|
-
`transformer_blocks.${layerIdx}.attn2.to_out.0.weight`,
|
|
490
|
-
numTokens,
|
|
491
|
-
hiddenSize,
|
|
492
|
-
hiddenSize,
|
|
493
|
-
recorder,
|
|
494
|
-
{ outputDtype: hiddenStates.dtype }
|
|
495
|
-
);
|
|
496
|
-
release(attention.buffer);
|
|
497
|
-
const outBias = createBiasTensor(getWeight(weightsEntry, `transformer_blocks.${layerIdx}.attn2.to_out.0.bias`), hiddenSize, 'sana_cross_attn_bias');
|
|
498
|
-
if (outBias) {
|
|
499
|
-
projected = await ops.biasAdd(projected, outBias, numTokens, hiddenSize);
|
|
500
|
-
}
|
|
501
|
-
return projected;
|
|
502
|
-
}
|
|
503
|
-
|
|
504
|
-
async function runGlumbConv(hiddenStates, layerIdx, weightsEntry, gridHeight, gridWidth, hiddenSize, recorder, runtime, ops, release) {
|
|
505
|
-
const expandRatio = Number(getWeightShape(weightsEntry, `transformer_blocks.${layerIdx}.ff.conv_inverted.weight`)[0]) / hiddenSize / 2;
|
|
506
|
-
const hiddenChannels = Math.floor(hiddenSize * expandRatio);
|
|
507
|
-
let inverted = await runMatmulResolved(
|
|
508
|
-
hiddenStates,
|
|
509
|
-
weightsEntry,
|
|
510
|
-
`transformer_blocks.${layerIdx}.ff.conv_inverted.weight`,
|
|
511
|
-
hiddenStates.shape[0],
|
|
512
|
-
hiddenChannels * 2,
|
|
513
|
-
hiddenSize,
|
|
514
|
-
recorder,
|
|
515
|
-
{ outputDtype: hiddenStates.dtype }
|
|
516
|
-
);
|
|
517
|
-
const invertedBias = createBiasTensor(getWeight(weightsEntry, `transformer_blocks.${layerIdx}.ff.conv_inverted.bias`), hiddenChannels * 2, 'sana_ff_inverted_bias');
|
|
518
|
-
if (invertedBias) {
|
|
519
|
-
inverted = await ops.biasAdd(inverted, invertedBias, hiddenStates.shape[0], hiddenChannels * 2);
|
|
520
|
-
}
|
|
521
|
-
const invertedAct = await ops.silu(inverted, { size: hiddenStates.shape[0] * hiddenChannels * 2, swigluLimit: null });
|
|
522
|
-
release(inverted.buffer);
|
|
523
|
-
|
|
524
|
-
const convInput = reshapeTensor(invertedAct, [hiddenChannels * 2, gridHeight, gridWidth], 'sana_ff_conv_input');
|
|
525
|
-
const depthWeight = expectDiffusionWeight(getWeight(weightsEntry, `transformer_blocks.${layerIdx}.ff.conv_depth.weight`), `transformer_blocks.${layerIdx}.ff.conv_depth.weight`);
|
|
526
|
-
const depthBias = getWeight(weightsEntry, `transformer_blocks.${layerIdx}.ff.conv_depth.bias`);
|
|
527
|
-
const depth = await ops.depthwiseConv2d(convInput, depthWeight, depthBias, {
|
|
528
|
-
channels: hiddenChannels * 2,
|
|
529
|
-
height: gridHeight,
|
|
530
|
-
width: gridWidth,
|
|
531
|
-
kernelH: 3,
|
|
532
|
-
kernelW: 3,
|
|
533
|
-
stride: 1,
|
|
534
|
-
pad: 1,
|
|
535
|
-
});
|
|
536
|
-
release(invertedAct.buffer);
|
|
537
|
-
|
|
538
|
-
const depthTokens = reshapeTensor(depth, [hiddenStates.shape[0], hiddenChannels * 2], 'sana_ff_depth_tokens');
|
|
539
|
-
const gated = await ops.siluRowSplit(depthTokens, {
|
|
540
|
-
numTokens: hiddenStates.shape[0],
|
|
541
|
-
dim: hiddenChannels,
|
|
542
|
-
activation: 'silu',
|
|
543
|
-
swigluLimit: null,
|
|
544
|
-
});
|
|
545
|
-
release(depth.buffer);
|
|
546
|
-
|
|
547
|
-
let projected = await runMatmulResolved(
|
|
548
|
-
gated,
|
|
549
|
-
weightsEntry,
|
|
550
|
-
`transformer_blocks.${layerIdx}.ff.conv_point.weight`,
|
|
551
|
-
hiddenStates.shape[0],
|
|
552
|
-
hiddenSize,
|
|
553
|
-
hiddenChannels,
|
|
554
|
-
recorder,
|
|
555
|
-
{ outputDtype: hiddenStates.dtype }
|
|
556
|
-
);
|
|
557
|
-
release(gated.buffer);
|
|
558
|
-
return projected;
|
|
559
|
-
}
|
|
560
|
-
|
|
561
|
-
export async function runSanaTransformer(latents, context, timeState, weightsEntry, modelConfig, runtime, options = {}) {
|
|
562
|
-
const device = getDevice();
|
|
563
|
-
if (!device) {
|
|
564
|
-
throw new Error('Sana transformer requires a WebGPU device.');
|
|
565
|
-
}
|
|
566
|
-
|
|
567
|
-
const recorder = options.recorder ?? null;
|
|
568
|
-
const ops = createKernelOps(recorder);
|
|
569
|
-
const release = createDiffusionBufferReleaser(recorder);
|
|
570
|
-
const destroy = createDiffusionBufferDestroyer(recorder);
|
|
571
|
-
const config = modelConfig?.components?.transformer?.config || {};
|
|
572
|
-
const hiddenSize = config.num_attention_heads * config.attention_head_dim;
|
|
573
|
-
const numHeads = config.num_attention_heads;
|
|
574
|
-
const headDim = config.attention_head_dim;
|
|
575
|
-
const patchSize = config.patch_size ?? 1;
|
|
576
|
-
const normEps = Number(config.norm_eps ?? 1e-6);
|
|
577
|
-
const latentHeight = latents.shape[1];
|
|
578
|
-
const latentWidth = latents.shape[2];
|
|
579
|
-
const gridHeight = Math.floor(latentHeight / patchSize);
|
|
580
|
-
const gridWidth = Math.floor(latentWidth / patchSize);
|
|
581
|
-
const numTokens = gridHeight * gridWidth;
|
|
582
|
-
|
|
583
|
-
const patchWeight = expectDiffusionWeight(getWeight(weightsEntry, 'patch_embed.proj.weight'), 'patch_embed.proj.weight');
|
|
584
|
-
const patchBias = getWeight(weightsEntry, 'patch_embed.proj.bias');
|
|
585
|
-
const conv = await ops.conv2d(latents, patchWeight, patchBias, {
|
|
586
|
-
inChannels: latents.shape[0],
|
|
587
|
-
outChannels: hiddenSize,
|
|
588
|
-
height: latentHeight,
|
|
589
|
-
width: latentWidth,
|
|
590
|
-
kernelH: patchSize,
|
|
591
|
-
kernelW: patchSize,
|
|
592
|
-
stride: patchSize,
|
|
593
|
-
pad: 0,
|
|
594
|
-
});
|
|
595
|
-
let hidden = await import('../../../gpu/kernels/transpose.js').then(({ runTranspose, recordTranspose }) => {
|
|
596
|
-
const transpose = recorder ? (input, rows, cols) => recordTranspose(recorder, input, rows, cols) : runTranspose;
|
|
597
|
-
return transpose(conv, hiddenSize, numTokens);
|
|
598
|
-
});
|
|
599
|
-
release(conv.buffer);
|
|
600
|
-
|
|
601
|
-
const ones = new Float32Array(hiddenSize).fill(1.0);
|
|
602
|
-
const zeros = new Float32Array(hiddenSize);
|
|
603
|
-
const onesBuf = createVectorBuffer(device, ones, 'sana_norm_ones');
|
|
604
|
-
const zerosBuf = createVectorBuffer(device, zeros, 'sana_norm_zeros');
|
|
605
|
-
|
|
606
|
-
for (let layerIdx = 0; layerIdx < config.num_layers; layerIdx++) {
|
|
607
|
-
const layerTable = expectDiffusionWeight(getWeight(weightsEntry, `transformer_blocks.${layerIdx}.scale_shift_table`), `transformer_blocks.${layerIdx}.scale_shift_table`);
|
|
608
|
-
const modulation = await buildLayerModulation(
|
|
609
|
-
timeState.modulation,
|
|
610
|
-
layerTable,
|
|
611
|
-
getWeightShape(weightsEntry, `transformer_blocks.${layerIdx}.scale_shift_table`),
|
|
612
|
-
recorder,
|
|
613
|
-
ops,
|
|
614
|
-
release
|
|
615
|
-
);
|
|
616
|
-
|
|
617
|
-
let norm1 = await ops.layerNorm(hidden, onesBuf, zerosBuf, normEps, { batchSize: numTokens, hiddenSize });
|
|
618
|
-
norm1 = await ops.modulate(norm1, modulation, {
|
|
619
|
-
numTokens,
|
|
620
|
-
hiddenSize,
|
|
621
|
-
scaleOffset: hiddenSize,
|
|
622
|
-
shiftOffset: 0,
|
|
623
|
-
gateOffset: hiddenSize * 2,
|
|
624
|
-
hasGate: false,
|
|
625
|
-
addOne: true,
|
|
626
|
-
});
|
|
627
|
-
const selfAttn = await runSelfAttention(norm1, layerIdx, weightsEntry, numTokens, hiddenSize, numHeads, headDim, normEps, recorder, runtime, ops, release);
|
|
628
|
-
release(norm1.buffer);
|
|
629
|
-
const gatedSelf = await ops.modulate(selfAttn, modulation, {
|
|
630
|
-
numTokens,
|
|
631
|
-
hiddenSize,
|
|
632
|
-
scaleOffset: hiddenSize * 2,
|
|
633
|
-
shiftOffset: hiddenSize * 6,
|
|
634
|
-
gateOffset: hiddenSize * 2,
|
|
635
|
-
hasGate: false,
|
|
636
|
-
addOne: false,
|
|
637
|
-
});
|
|
638
|
-
release(selfAttn.buffer);
|
|
639
|
-
let nextHidden = await ops.residualAdd(hidden, gatedSelf, numTokens * hiddenSize, { useVec4: true });
|
|
640
|
-
release(hidden.buffer);
|
|
641
|
-
release(gatedSelf.buffer);
|
|
642
|
-
hidden = createTensor(nextHidden.buffer, nextHidden.dtype, [numTokens, hiddenSize], 'sana_hidden_after_self');
|
|
643
|
-
|
|
644
|
-
let norm2 = await ops.layerNorm(hidden, onesBuf, zerosBuf, normEps, { batchSize: numTokens, hiddenSize });
|
|
645
|
-
const crossAttn = await runCrossAttention(norm2, context, layerIdx, weightsEntry, numTokens, hiddenSize, config, recorder, ops, release);
|
|
646
|
-
release(norm2.buffer);
|
|
647
|
-
nextHidden = await ops.residualAdd(hidden, crossAttn, numTokens * hiddenSize, { useVec4: true });
|
|
648
|
-
release(hidden.buffer);
|
|
649
|
-
release(crossAttn.buffer);
|
|
650
|
-
hidden = createTensor(nextHidden.buffer, nextHidden.dtype, [numTokens, hiddenSize], 'sana_hidden_after_cross');
|
|
651
|
-
|
|
652
|
-
let normFf = await ops.layerNorm(hidden, onesBuf, zerosBuf, normEps, { batchSize: numTokens, hiddenSize });
|
|
653
|
-
normFf = await ops.modulate(normFf, modulation, {
|
|
654
|
-
numTokens,
|
|
655
|
-
hiddenSize,
|
|
656
|
-
scaleOffset: hiddenSize * 4,
|
|
657
|
-
shiftOffset: hiddenSize * 3,
|
|
658
|
-
gateOffset: hiddenSize * 5,
|
|
659
|
-
hasGate: false,
|
|
660
|
-
addOne: true,
|
|
661
|
-
});
|
|
662
|
-
const ff = await runGlumbConv(normFf, layerIdx, weightsEntry, gridHeight, gridWidth, hiddenSize, recorder, runtime, ops, release);
|
|
663
|
-
release(normFf.buffer);
|
|
664
|
-
const gatedFf = await ops.modulate(ff, modulation, {
|
|
665
|
-
numTokens,
|
|
666
|
-
hiddenSize,
|
|
667
|
-
scaleOffset: hiddenSize * 5,
|
|
668
|
-
shiftOffset: hiddenSize * 6,
|
|
669
|
-
gateOffset: hiddenSize * 5,
|
|
670
|
-
hasGate: false,
|
|
671
|
-
addOne: false,
|
|
672
|
-
});
|
|
673
|
-
release(ff.buffer);
|
|
674
|
-
release(modulation.buffer);
|
|
675
|
-
nextHidden = await ops.residualAdd(hidden, gatedFf, numTokens * hiddenSize, { useVec4: true });
|
|
676
|
-
release(hidden.buffer);
|
|
677
|
-
release(gatedFf.buffer);
|
|
678
|
-
hidden = createTensor(nextHidden.buffer, nextHidden.dtype, [numTokens, hiddenSize], 'sana_hidden_after_ff');
|
|
679
|
-
}
|
|
680
|
-
release(timeState.modulation.buffer);
|
|
681
|
-
|
|
682
|
-
const finalTable = expectDiffusionWeight(getWeight(weightsEntry, 'scale_shift_table'), 'scale_shift_table');
|
|
683
|
-
const duplicated = await duplicateVectorTensor(timeState.embeddedTimestep, 2, recorder);
|
|
684
|
-
release(timeState.embeddedTimestep.buffer);
|
|
685
|
-
let finalMod = await ops.biasAdd(
|
|
686
|
-
duplicated,
|
|
687
|
-
createTensor(getBuffer(finalTable), duplicated.dtype, [hiddenSize * 2], 'sana_final_table'),
|
|
688
|
-
1,
|
|
689
|
-
hiddenSize * 2
|
|
690
|
-
);
|
|
691
|
-
release(duplicated.buffer);
|
|
692
|
-
let normed = await ops.layerNorm(hidden, onesBuf, zerosBuf, 1e-6, { batchSize: numTokens, hiddenSize });
|
|
693
|
-
normed = await ops.modulate(normed, finalMod, {
|
|
694
|
-
numTokens,
|
|
695
|
-
hiddenSize,
|
|
696
|
-
scaleOffset: hiddenSize,
|
|
697
|
-
shiftOffset: 0,
|
|
698
|
-
gateOffset: hiddenSize,
|
|
699
|
-
hasGate: false,
|
|
700
|
-
addOne: true,
|
|
701
|
-
});
|
|
702
|
-
release(hidden.buffer);
|
|
703
|
-
release(finalMod.buffer);
|
|
704
|
-
|
|
705
|
-
let projected = await runMatmulResolved(
|
|
706
|
-
normed,
|
|
707
|
-
weightsEntry,
|
|
708
|
-
'proj_out.weight',
|
|
709
|
-
numTokens,
|
|
710
|
-
config.out_channels ?? latents.shape[0],
|
|
711
|
-
hiddenSize,
|
|
712
|
-
recorder,
|
|
713
|
-
{ outputDtype: normed.dtype }
|
|
714
|
-
);
|
|
715
|
-
release(normed.buffer);
|
|
716
|
-
const projBias = createBiasTensor(getWeight(weightsEntry, 'proj_out.bias'), projected.shape[1], 'sana_proj_out_bias');
|
|
717
|
-
if (projBias) {
|
|
718
|
-
projected = await ops.biasAdd(projected, projBias, numTokens, projected.shape[1]);
|
|
719
|
-
}
|
|
720
|
-
|
|
721
|
-
const { runTranspose, recordTranspose } = await import('../../../gpu/kernels/transpose.js');
|
|
722
|
-
const transpose = recorder ? (input, rows, cols) => recordTranspose(recorder, input, rows, cols) : runTranspose;
|
|
723
|
-
const channelsFirst = await transpose(projected, numTokens, projected.shape[1]);
|
|
724
|
-
release(projected.buffer);
|
|
725
|
-
destroy(onesBuf);
|
|
726
|
-
destroy(zerosBuf);
|
|
727
|
-
return reshapeTensor(channelsFirst, [config.out_channels ?? latents.shape[0], gridHeight, gridWidth], 'sana_output');
|
|
728
|
-
}
|
|
729
|
-
|
|
730
|
-
export async function buildSanaConditioning(context, attentionMask, timestep, guidanceScale, weightsEntry, modelConfig, runtime, options = {}) {
|
|
731
|
-
const config = modelConfig?.components?.transformer?.config || {};
|
|
732
|
-
const projectedContext = await projectSanaContext(context, attentionMask, weightsEntry, config, runtime, options);
|
|
733
|
-
const timestepState = await buildSanaTimestepConditioning(timestep, guidanceScale, weightsEntry, config, runtime, options);
|
|
734
|
-
return {
|
|
735
|
-
context: projectedContext,
|
|
736
|
-
timeState: timestepState,
|
|
737
|
-
};
|
|
738
|
-
}
|