doppler-gpu 0.4.6 → 0.4.8
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/CHANGELOG.md +64 -0
- package/README.md +98 -137
- package/assets/doppler.svg +1 -3
- package/models/catalog.json +1575 -0
- package/package.json +59 -26
- package/src/cli/doppler-cli.js +28 -7
- package/src/cli/doppler-quickstart.js +1 -1
- package/src/cli/doppler-serve.js +2 -2
- package/src/client/doppler-registry.js +12 -0
- package/src/client/doppler-registry.json +52 -3
- package/src/client/runtime/index.js +25 -2
- package/src/client/runtime/model-session.d.ts +31 -1
- package/src/client/runtime/model-session.js +31 -0
- package/src/client/runtime/model-source.d.ts +19 -2
- package/src/client/runtime/model-source.js +45 -5
- package/src/config/conversion/diffusiongemma/diffusiongemma-26b-a4b-it-q4k-ehf16-af16.json +28 -4
- package/src/config/conversion/embeddinggemma/google-embeddinggemma-300m-q4k-ehf16-af32.json +29 -4
- package/src/config/conversion/gemma3/gemma-3-1b-it-f16-af32.json +29 -4
- package/src/config/conversion/gemma3/gemma-3-1b-it-q4k-ehf16-af32.json +29 -4
- package/src/config/conversion/gemma3/gemma-3-270m-it-f16-af32.json +29 -4
- package/src/config/conversion/gemma3/gemma-3-270m-it-q4k-ehf16-af16.json +29 -4
- package/src/config/conversion/gemma3/gemma-3-270m-it-q4k-ehf16-af32.json +29 -4
- package/src/config/conversion/gemma3/translategemma-4b-1b-enes-q4k-ehf16-af32.json +50 -9
- package/src/config/conversion/gemma3/translategemma-4b-it-q4k-ehf16-af32.json +35 -5
- package/src/config/conversion/gemma4/gemma-4-12b-it-text-q4k-ehf16-af16.json +31 -5
- package/src/config/conversion/gemma4/gemma-4-12b-it-text-q4k-ehf16-af32.json +30 -4
- package/src/config/conversion/gemma4/gemma-4-12b-it-text-q4k-ehf16-hq4k-af16.json +30 -4
- package/src/config/conversion/gemma4/gemma-4-12b-it-text-w4a16-ct-ehf16-af16.json +30 -4
- package/src/config/conversion/gemma4/gemma-4-31b-it-text-q4k-ehf16-af16.json +30 -5
- package/src/config/conversion/gemma4/gemma-4-31b-it-text-q4k-ehf16-af32.json +29 -4
- package/src/config/conversion/gemma4/gemma-4-e2b-it-q4k-ehf16-af16-int4ple.json +30 -5
- package/src/config/conversion/gemma4/gemma-4-e2b-it-q4k-ehf16-af32-int4ple.json +29 -4
- package/src/config/conversion/gemma4/gemma-4-e2b-it-q4k-ehf16-af32.json +29 -4
- package/src/config/conversion/gemma4/gemma-4-moe-q4k-ehf16-af32.json +29 -4
- package/src/config/conversion/gpt-oss-20b-f16-xmxfp4.json +29 -4
- package/src/config/conversion/janus/janus-pro-1b-text-q4k-ehaf16.json +29 -4
- package/src/config/conversion/lfm2/lfm2.5-1.2b-instruct-q4k-ehf16-af32.json +29 -4
- package/src/config/conversion/minicpm/minicpm4-0-5b-f16-af32.json +481 -0
- package/src/config/conversion/qwen3/qwen-3-5-0-8b-q4k-ehaf16.json +29 -4
- package/src/config/conversion/qwen3/qwen-3-5-2b-q4k-ehaf16.json +29 -4
- package/src/config/conversion/qwen3/qwen-3-6-27b-q4k-eaf16.json +30 -5
- package/src/config/conversion/qwen3/qwen-3-6-27b-q4k-ehaf16.json +29 -4
- package/src/config/conversion/qwen3/qwen-3-embedding-0-6b-q4k-ehf16-af32.json +395 -0
- package/src/config/conversion/qwen3/qwen-3-reranker-0-6b-f16-af32.json +429 -0
- package/src/config/conversion/qwen3/qwen-3-reranker-0-6b-q4k-ehf16-af32.json +548 -0
- package/src/config/execution-contract-check.js +14 -10
- package/src/config/kernel-path-loader.js +1 -0
- package/src/config/kernels/kernel-ref-digests.js +26 -17
- package/src/config/kernels/registry.json +448 -27
- package/src/config/merge.js +41 -0
- package/src/config/param-validator.js +67 -8
- package/src/config/required-inference-fields-contract-check.js +14 -0
- package/src/config/runtime/default.json +3 -1
- package/src/config/runtime/profiles/qwen-3-5-0-8b-metal-attn128-candidate.json +20 -0
- package/src/config/runtime/profiles/qwen-3-5-0-8b-metal-maxbatch64-candidate.json +52 -0
- package/src/config/runtime/profiles/qwen-3-5-0-8b-metal-normed-ffn-candidate.json +18 -0
- package/src/config/runtime/profiles/qwen-3-5-0-8b-metal-overlap-ringtokens1-candidate.json +52 -0
- package/src/config/runtime/profiles/qwen-3-5-0-8b-metal-parity.json +85 -0
- package/src/config/runtime/profiles/qwen-3-5-0-8b-metal-sequential-candidate.json +52 -0
- package/src/config/runtime/profiles/qwen-3-5-0-8b-metal-throughput.json +85 -0
- package/src/config/runtime/profiles/qwen-3-5-2b-f16-primary-throughput.json +23 -0
- package/src/config/runtime/profiles/qwen-3-5-2b-metal-parity.json +38 -0
- package/src/config/runtime/profiles/qwen-3-5-2b-metal-throughput.json +135 -0
- package/src/config/runtime/profiles/qwen-3-embedding-0-6b-metal-stability.json +37 -0
- package/src/config/runtime/profiles/qwen-3-embedding-0-6b-metal-throughput.json +21 -0
- package/src/config/runtime/profiles/qwen-3-reranker-0-6b-metal-stability.json +41 -0
- package/src/config/runtime/profiles/qwen-3-reranker-0-6b-metal-throughput.json +41 -0
- package/src/config/runtime/profiles/rerank-stability.json +43 -0
- package/src/config/runtime/profiles/rerank-throughput.json +40 -0
- package/src/config/runtime/profiles/translategemma-1b-nativekd2-q4k-rdna3-throughput-probe.json +22 -0
- package/src/config/schema/debug.schema.d.ts +2 -0
- package/src/config/schema/debug.schema.js +1 -0
- package/src/config/schema/doppler.schema.d.ts +5 -1
- package/src/config/schema/doppler.schema.js +117 -292
- package/src/config/schema/execution-v1.schema.d.ts +41 -0
- package/src/config/schema/execution-v1.schema.js +10 -0
- package/src/config/schema/index.d.ts +3 -1
- package/src/config/schema/inference.schema.d.ts +17 -3
- package/src/config/schema/kvcache.schema.d.ts +19 -3
- package/src/config/schema/kvcache.schema.js +2 -1
- package/src/config/schema/manifest.schema.d.ts +38 -1
- package/src/config/schema/manifest.schema.js +18 -0
- package/src/config/schema/speculation-self.schema.d.ts +16 -0
- package/src/config/source-packages/litert/gemma-4-12b-it.json +20 -1
- package/src/config/source-packages/litert/gemma-4-e2b-it.json +20 -1
- package/src/config/source-packages/litert/gemma-4-e4b-it.json +21 -1
- package/src/config/transforms/capability-transform-resolver.js +7 -1
- package/src/config/transforms/execution-graph-transforms.js +10 -2
- package/src/converter/conversion-plan.js +2 -0
- package/src/converter/parsers/transformer.js +27 -1
- package/src/converter/quantization-info.js +16 -4
- package/src/converter/rope-config.js +55 -3
- package/src/debug/config.d.ts +4 -0
- package/src/experimental/orchestration/index.d.ts +10 -1
- package/src/experimental/training/distillation/student-fixture.js +1 -1
- package/src/experimental/training/lora-pipeline.js +1 -1
- package/src/experimental/training/lora.js +2 -2
- package/src/experimental/training/suite.js +2 -2
- package/src/formats/rdrr/classification.js +3 -0
- package/src/formats/rdrr/integrity.d.ts +0 -2
- package/src/formats/rdrr/integrity.js +0 -2
- package/src/formats/rdrr/types.d.ts +1 -1
- package/src/gpu/kernels/attention.d.ts +2 -0
- package/src/gpu/kernels/attention.js +75 -16
- package/src/gpu/kernels/attention_decode_online_head256_f16kv_output_gate.wgsl +268 -0
- package/src/gpu/kernels/attention_head128_f16kv.wgsl +269 -0
- package/src/gpu/kernels/backward/attention_backward.js +8 -0
- package/src/gpu/kernels/dequant.d.ts +6 -6
- package/src/gpu/kernels/dequant.js +17 -6
- package/src/gpu/kernels/fused_ffn.d.ts +32 -0
- package/src/gpu/kernels/fused_ffn.js +540 -19
- package/src/gpu/kernels/fused_ffn.wgsl +2 -1
- package/src/gpu/kernels/fused_ffn_f16.wgsl +2 -1
- package/src/gpu/kernels/fused_ffn_f16w_f32a_batched.wgsl +135 -0
- package/src/gpu/kernels/fused_ffn_q4k.wgsl +5 -3
- package/src/gpu/kernels/fused_ffn_q4k_f16.wgsl +2 -1
- package/src/gpu/kernels/fused_ffn_q4k_metal_simd16.wgsl +183 -0
- package/src/gpu/kernels/fused_gate_up_gelu_gated_f16.wgsl +2 -1
- package/src/gpu/kernels/fused_matmul_q4.wgsl +15 -6
- package/src/gpu/kernels/fused_matmul_q4_metal_simd16.wgsl +144 -0
- package/src/gpu/kernels/index.d.ts +21 -0
- package/src/gpu/kernels/index.js +13 -0
- package/src/gpu/kernels/linear-attention-core.js +414 -21
- package/src/gpu/kernels/lm-head-argmax.d.ts +10 -1
- package/src/gpu/kernels/lm-head-argmax.js +153 -57
- package/src/gpu/kernels/lm-head-select-logits.d.ts +23 -0
- package/src/gpu/kernels/lm-head-select-logits.js +186 -0
- package/src/gpu/kernels/lm_head_argmax_q4k.wgsl +289 -0
- package/src/gpu/kernels/lm_head_select_logits.wgsl +77 -0
- package/src/gpu/kernels/matmul-selection.js +9 -2
- package/src/gpu/kernels/matmul.d.ts +1 -1
- package/src/gpu/kernels/matmul.js +24 -6
- package/src/gpu/kernels/moe.d.ts +5 -1
- package/src/gpu/kernels/moe.js +5 -2
- package/src/gpu/kernels/residual_rmsnorm_pair.wgsl +127 -0
- package/src/gpu/kernels/rmsnorm-pair.d.ts +33 -0
- package/src/gpu/kernels/rmsnorm-pair.js +113 -0
- package/src/gpu/kernels/rmsnorm-stats.d.ts +29 -0
- package/src/gpu/kernels/rmsnorm-stats.js +250 -0
- package/src/gpu/kernels/silu.wgsl +7 -1
- package/src/gpu/kernels/silu_f16.wgsl +7 -1
- package/src/gpu/kernels/softmax.d.ts +3 -3
- package/src/gpu/kernels/softmax.js +9 -2
- package/src/gpu/kernels/split-qkv-rmsnorm-rope-qk.js +15 -10
- package/src/gpu/kernels/split_qkv_rmsnorm_rope_qk.wgsl +37 -7
- package/src/gpu/kernels/split_qkv_rmsnorm_rope_qk_f16kv_cache.wgsl +37 -7
- package/src/gpu/kernels/turboquant-codebook.js +1 -1
- package/src/gpu/kernels/types.d.ts +1 -1
- package/src/gpu/profiler.js +11 -1
- package/src/gpu/tensor.d.ts +1 -0
- package/src/index.d.ts +1 -1
- package/src/index.js +1 -1
- package/src/inference/browser-harness-model-helpers.js +4 -1
- package/src/inference/browser-harness-suite-helpers.js +1 -1
- package/src/inference/browser-harness-text-helpers.js +589 -117
- package/src/inference/browser-harness.d.ts +3 -3
- package/src/inference/browser-harness.js +486 -13
- package/src/inference/fixtures/rerank-semantic-fixtures.json +38 -0
- package/src/inference/kv-cache/base.js +42 -1
- package/src/inference/kv-cache/quantized.js +16 -2
- package/src/inference/kv-cache/tiered.js +50 -15
- package/src/inference/moe-router.js +38 -7
- package/src/inference/pipelines/diffusion/text-encoder-gpu.js +8 -6
- package/src/inference/pipelines/structured/json-grammar-mask.d.ts +18 -0
- package/src/inference/pipelines/text/attention/dispatch-params.js +8 -3
- package/src/inference/pipelines/text/attention/output-gate-fusion.js +37 -0
- package/src/inference/pipelines/text/attention/projections.js +174 -4
- package/src/inference/pipelines/text/attention/record.js +44 -21
- package/src/inference/pipelines/text/attention/run.js +18 -3
- package/src/inference/pipelines/text/config.d.ts +23 -1
- package/src/inference/pipelines/text/config.js +235 -112
- package/src/inference/pipelines/text/embed.js +39 -12
- package/src/inference/pipelines/text/execution-plan.d.ts +2 -0
- package/src/inference/pipelines/text/execution-plan.js +45 -2
- package/src/inference/pipelines/text/execution-v1.d.ts +6 -0
- package/src/inference/pipelines/text/execution-v1.js +122 -8
- package/src/inference/pipelines/text/ffn/dense.js +254 -25
- package/src/inference/pipelines/text/ffn/moe.js +4 -3
- package/src/inference/pipelines/text/ffn/standard.js +145 -39
- package/src/inference/pipelines/text/generator-decode-policy.js +1 -0
- package/src/inference/pipelines/text/generator-helpers.js +4 -1
- package/src/inference/pipelines/text/generator-runtime.js +54 -20
- package/src/inference/pipelines/text/generator-steps.js +5 -2
- package/src/inference/pipelines/text/generator.d.ts +18 -1
- package/src/inference/pipelines/text/generator.js +303 -20
- package/src/inference/pipelines/text/init-qkv-fusion.js +155 -62
- package/src/inference/pipelines/text/init-rope.js +54 -6
- package/src/inference/pipelines/text/init.d.ts +22 -3
- package/src/inference/pipelines/text/init.js +65 -17
- package/src/inference/pipelines/text/layer.js +111 -5
- package/src/inference/pipelines/text/linear-attention-ab-fusion.js +208 -0
- package/src/inference/pipelines/text/linear-attention.d.ts +1 -0
- package/src/inference/pipelines/text/linear-attention.js +185 -24
- package/src/inference/pipelines/text/logits/gpu.d.ts +1 -1
- package/src/inference/pipelines/text/logits/gpu.js +78 -115
- package/src/inference/pipelines/text/logits/index.js +105 -87
- package/src/inference/pipelines/text/logits/precision-policy.js +103 -0
- package/src/inference/pipelines/text/logits/types.d.ts +1 -0
- package/src/inference/pipelines/text/logits/utils.d.ts +2 -0
- package/src/inference/pipelines/text/logits/utils.js +11 -0
- package/src/inference/pipelines/text/lora-apply.js +6 -4
- package/src/inference/pipelines/text/model-load.js +6 -0
- package/src/inference/pipelines/text/moe-cpu.js +25 -0
- package/src/inference/pipelines/text/moe-gpu.js +148 -97
- package/src/inference/pipelines/text/moe-shape-validator.js +149 -31
- package/src/inference/pipelines/text/ops.js +48 -1
- package/src/inference/pipelines/text/per-layer-inputs.js +105 -21
- package/src/inference/pipelines/text/probes.js +9 -4
- package/src/inference/pipelines/text/state.d.ts +4 -4
- package/src/inference/pipelines/text/state.js +4 -0
- package/src/inference/pipelines/text/types.d.ts +27 -2
- package/src/inference/pipelines/text/weights.js +7 -2
- package/src/inference/pipelines/text.d.ts +21 -1
- package/src/inference/pipelines/text.js +60 -14
- package/src/inference/tokenizers/behavior-flags.js +23 -2
- package/src/inference/tokenizers/bundled.js +45 -23
- package/src/inference/tokenizers/types.d.ts +7 -4
- package/src/loader/doppler-loader.js +25 -2
- package/src/loader/final-weights-loader.js +12 -1
- package/src/loader/index.d.ts +3 -1
- package/src/loader/index.js +2 -0
- package/src/loader/loader-types.d.ts +3 -1
- package/src/loader/manifest-config.d.ts +18 -0
- package/src/loader/manifest-config.js +11 -3
- package/src/loader/shard-cache.d.ts +2 -0
- package/src/loader/shard-cache.js +29 -1
- package/src/models/gemma4.js +8 -0
- package/src/rules/execution-rules-contract-check.js +12 -3
- package/src/rules/inference/capability-transforms.rules.json +159 -22
- package/src/rules/inference/config.rules.json +26 -2
- package/src/rules/inference/dtype.rules.json +12 -5
- package/src/rules/inference/execution.rules.json +5 -19
- package/src/rules/inference/ffn.rules.json +33 -0
- package/src/rules/inference/moe.rules.json +97 -0
- package/src/rules/kernels/fused-ffn.rules.json +1 -1
- package/src/rules/kernels/lm-head-argmax.rules.json +10 -0
- package/src/rules/kernels/matmul.rules.json +1 -0
- package/src/rules/kernels/silu.rules.json +1 -0
- package/src/rules/rule-registry.js +2 -0
- package/src/storage/artifact-storage-context.js +36 -0
- package/src/tooling/command-api-constants.js +2 -2
- package/src/tooling/command-api-family-normalizers.js +2 -1
- package/src/tooling/command-api.d.ts +1 -1
- package/src/tooling/command-runner-shared.js +3 -1
- package/src/tooling/node-browser-command-runner.js +42 -7
- package/src/tooling/node-converter.js +20 -0
- package/src/tooling/node-source-runtime.js +5 -1
- package/src/tooling/opfs-cache.d.ts +21 -1
- package/src/tooling/opfs-cache.js +168 -32
- package/src/tooling-exports/storage.d.ts +1 -1
- package/src/tooling-exports/storage.js +1 -1
- package/src/types/gpu.d.ts +25 -0
- package/src/version.js +1 -1
- package/tests/kernels/browser/kernel-suite.js +1593 -0
- package/tests/kernels/browser/test-page.js +2629 -0
- package/tests/kernels/harness/benchmark.js +225 -0
- package/tests/kernels/harness/buffer-utils.js +88 -0
- package/tests/kernels/harness/tolerance.js +174 -0
- package/tests/kernels/reference/adam.js +15 -0
- package/tests/kernels/reference/attention.js +158 -0
- package/tests/kernels/reference/backward.js +96 -0
- package/tests/kernels/reference/bias-add.js +20 -0
- package/tests/kernels/reference/conv2d.js +58 -0
- package/tests/kernels/reference/cross-entropy.js +26 -0
- package/tests/kernels/reference/dequant.js +312 -0
- package/tests/kernels/reference/embed.js +15 -0
- package/tests/kernels/reference/gather.js +61 -0
- package/tests/kernels/reference/gelu.js +37 -0
- package/tests/kernels/reference/groupnorm.js +36 -0
- package/tests/kernels/reference/index.js +89 -0
- package/tests/kernels/reference/layernorm.js +22 -0
- package/tests/kernels/reference/matmul.js +57 -0
- package/tests/kernels/reference/modulate.js +24 -0
- package/tests/kernels/reference/moe-gather.js +83 -0
- package/tests/kernels/reference/pixel-shuffle.js +20 -0
- package/tests/kernels/reference/residual.js +55 -0
- package/tests/kernels/reference/rmsnorm.js +52 -0
- package/tests/kernels/reference/rope.js +79 -0
- package/tests/kernels/reference/sample.js +97 -0
- package/tests/kernels/reference/scatter-add.js +58 -0
- package/tests/kernels/reference/silu.js +52 -0
- package/tests/kernels/reference/softmax.js +86 -0
- package/tests/kernels/reference/split-qkv.js +58 -0
- package/tests/kernels/reference/topk.js +88 -0
- package/tests/kernels/reference/upsample2d.js +37 -0
- package/BRANDING.md +0 -14
- package/SECURITY.md +0 -19
- package/assets/doppler-readme-hero.svg +0 -84
- package/assets/doppler-runtime-map.svg +0 -84
- package/src/bootstrap.js +0 -30
- package/src/cli/doppler-cli.d.ts +0 -46
- package/src/cli/doppler-quickstart.d.ts +0 -45
- package/src/cli/doppler-serve.d.ts +0 -210
- package/src/cli/runtime-profiles.d.ts +0 -40
- package/src/client/doppler-api.browser.d.ts +0 -1
- package/src/client/doppler-api.browser.js +0 -25
- package/src/client/doppler-registry.d.ts +0 -35
- package/src/client/runtime/node-quickstart-cache.d.ts +0 -22
- package/src/client/runtime/source-runtime.d.ts +0 -26
- package/src/client/runtime/storage.d.ts +0 -5
- package/src/config/README.md +0 -61
- package/src/config/conversion/README.md +0 -58
- package/src/config/goal-completion-matrix.json +0 -552
- package/src/config/kernels/kernel-ref-digests.d.ts +0 -1
- package/src/config/kernels/moe/gpt-oss.paths.json +0 -49
- package/src/config/kernels/moe/mixtral.paths.json +0 -46
- package/src/config/kernels/registry.d.ts +0 -51
- package/src/config/merge-contract-check.d.ts +0 -16
- package/src/config/merge-contract-check.js +0 -318
- package/src/config/merge-helpers.d.ts +0 -60
- package/src/config/model-lanes/catalog-lane-resolver.d.ts +0 -66
- package/src/config/model-lanes/catalog-lane-resolver.js +0 -170
- package/src/config/param-categories.d.ts +0 -8
- package/src/config/param-validator.d.ts +0 -28
- package/src/config/platforms/loader.d.ts +0 -38
- package/src/config/quantization-contract-check.d.ts +0 -12
- package/src/config/quantization-contract-check.js +0 -91
- package/src/config/runtime/README.md +0 -28
- package/src/config/runtime-merge.d.ts +0 -5
- package/src/config/schema/claim-evidence-contract.schema.json +0 -93
- package/src/config/schema/command-surface-contract.schema.json +0 -81
- package/src/config/schema/goal-completion-matrix.schema.json +0 -223
- package/src/config/schema/model-artifact-contract.schema.json +0 -119
- package/src/config/schema/platform.schema.js +0 -1
- package/src/config/schema/policy-schema-registry.json +0 -26
- package/src/config/schema/subsystem-support-contract.schema.json +0 -48
- package/src/config/support-tiers/subsystems.json +0 -502
- package/src/config/supported-operations.d.ts +0 -14
- package/src/config/transforms/capability-transform-resolver.d.ts +0 -56
- package/src/config/transforms/execution-graph-transforms.d.ts +0 -300
- package/src/converter/conversion-plan.d.ts +0 -56
- package/src/converter/index.d.ts +0 -98
- package/src/converter/index.js +0 -59
- package/src/converter/parsers/diffusion.d.ts +0 -50
- package/src/converter/parsers/gguf.d.ts +0 -31
- package/src/converter/parsers/index.d.ts +0 -21
- package/src/converter/parsers/index.js +0 -12
- package/src/converter/parsers/transformer.d.ts +0 -31
- package/src/converter/quantization-info.d.ts +0 -38
- package/src/converter/rope-config.d.ts +0 -16
- package/src/converter/tokenizer-utils.d.ts +0 -12
- package/src/debug/reference/README.md +0 -114
- package/src/debug/reference/hf_attn_debug.py +0 -114
- package/src/debug/reference/hf_embed_check.py +0 -89
- package/src/debug/reference/hf_layer_out.py +0 -100
- package/src/debug/reference/hf_qwen35_linear_attn_debug.py +0 -268
- package/src/debug/reference/hf_rope_check.py +0 -116
- package/src/debug/reference/hf_weights.py +0 -75
- package/src/errors/doppler-error.d.ts +0 -20
- package/src/experimental/bridge/extension/background.d.ts +0 -14
- package/src/experimental/bridge/extension/background.js +0 -168
- package/src/experimental/bridge/extension/manifest.json +0 -34
- package/src/experimental/bridge/extension-client.d.ts +0 -114
- package/src/experimental/bridge/index.d.ts +0 -69
- package/src/experimental/bridge/node-bridge.js +0 -24
- package/src/experimental/bridge/protocol.d.ts +0 -96
- package/src/experimental/browser/gguf-importer.d.ts +0 -136
- package/src/experimental/browser/gguf-importer.js +0 -532
- package/src/experimental/browser/gguf-parser-browser.d.ts +0 -14
- package/src/experimental/browser/quantization.d.ts +0 -74
- package/src/experimental/browser/safetensors-parser-browser.d.ts +0 -193
- package/src/experimental/browser/shard-io-browser.d.ts +0 -57
- package/src/experimental/browser/shard-io-browser.js +0 -89
- package/src/experimental/browser/tensor-source-download.d.ts +0 -27
- package/src/experimental/diffusion/config/expansion-lanes.json +0 -29
- package/src/experimental/distribution/p2p-transport-contract.d.ts +0 -57
- package/src/experimental/distribution/p2p-transport.js +0 -75
- package/src/experimental/hotswap/intent-bundle.d.ts +0 -37
- package/src/experimental/hotswap/manifest.d.ts +0 -42
- package/src/experimental/hotswap/model-swap.js +0 -42
- package/src/experimental/hotswap/runtime.d.ts +0 -31
- package/src/experimental/training/README.md +0 -164
- package/src/experimental/training/artifacts.d.ts +0 -160
- package/src/experimental/training/attention-backward.d.ts +0 -30
- package/src/experimental/training/attention-forward.d.ts +0 -22
- package/src/experimental/training/attention-forward.js +0 -82
- package/src/experimental/training/autograd.d.ts +0 -51
- package/src/experimental/training/backward-kernels.js +0 -39
- package/src/experimental/training/checkpoint-watch.d.ts +0 -8
- package/src/experimental/training/checkpoint.d.ts +0 -36
- package/src/experimental/training/clip.d.ts +0 -9
- package/src/experimental/training/dataloader.d.ts +0 -8
- package/src/experimental/training/datasets/gepa-frontier.d.ts +0 -41
- package/src/experimental/training/datasets/gepa-frontier.js +0 -220
- package/src/experimental/training/datasets/index.d.ts +0 -24
- package/src/experimental/training/datasets/index.js +0 -13
- package/src/experimental/training/datasets/jsonl.d.ts +0 -11
- package/src/experimental/training/datasets/reploid.d.ts +0 -3
- package/src/experimental/training/datasets/reploid.js +0 -36
- package/src/experimental/training/datasets/teacher-traces.d.ts +0 -51
- package/src/experimental/training/datasets/teacher-traces.js +0 -279
- package/src/experimental/training/datasets/text-pairs.d.ts +0 -66
- package/src/experimental/training/datasets/token-batch.d.ts +0 -21
- package/src/experimental/training/datasets/token-batch.js +0 -52
- package/src/experimental/training/datasets/translation-pairs.d.ts +0 -34
- package/src/experimental/training/datasets/translation-pairs.js +0 -49
- package/src/experimental/training/distillation/artifacts.d.ts +0 -71
- package/src/experimental/training/distillation/checkpoint-watch.d.ts +0 -10
- package/src/experimental/training/distillation/dataset.d.ts +0 -59
- package/src/experimental/training/distillation/eval.d.ts +0 -34
- package/src/experimental/training/distillation/index.d.ts +0 -29
- package/src/experimental/training/distillation/runtime.d.ts +0 -20
- package/src/experimental/training/distillation/scoreboard.d.ts +0 -6
- package/src/experimental/training/distillation/stage-a.d.ts +0 -45
- package/src/experimental/training/distillation/stage-b.d.ts +0 -24
- package/src/experimental/training/distillation/student-fixture.d.ts +0 -27
- package/src/experimental/training/distillation/suite-data.d.ts +0 -45
- package/src/experimental/training/export.d.ts +0 -41
- package/src/experimental/training/index.d.ts +0 -74
- package/src/experimental/training/index.js +0 -63
- package/src/experimental/training/lora-pipeline.d.ts +0 -183
- package/src/experimental/training/lora.d.ts +0 -19
- package/src/experimental/training/loss-scaling.d.ts +0 -21
- package/src/experimental/training/loss.d.ts +0 -10
- package/src/experimental/training/objectives/base.d.ts +0 -58
- package/src/experimental/training/objectives/cross_entropy.d.ts +0 -18
- package/src/experimental/training/objectives/distill_kd.d.ts +0 -16
- package/src/experimental/training/objectives/distill_triplet.d.ts +0 -16
- package/src/experimental/training/objectives/index.d.ts +0 -12
- package/src/experimental/training/objectives/index.js +0 -6
- package/src/experimental/training/objectives/ul_stage1_joint.d.ts +0 -16
- package/src/experimental/training/objectives/ul_stage2_base.d.ts +0 -16
- package/src/experimental/training/operator-agent-eval.d.ts +0 -65
- package/src/experimental/training/operator-artifacts.d.ts +0 -62
- package/src/experimental/training/operator-command.d.ts +0 -5
- package/src/experimental/training/operator-eval.d.ts +0 -48
- package/src/experimental/training/operator-scoreboard.d.ts +0 -5
- package/src/experimental/training/optimizer.d.ts +0 -22
- package/src/experimental/training/runner.d.ts +0 -249
- package/src/experimental/training/suite.d.ts +0 -299
- package/src/experimental/training/tensor-factory.d.ts +0 -9
- package/src/experimental/training/trainer.d.ts +0 -89
- package/src/experimental/training/ul_dataset.d.ts +0 -47
- package/src/experimental/training/ul_schedule.d.ts +0 -6
- package/src/experimental/training/workload-packs/distill-glm52-doppler-agent-harness-gemma270m-f16-sft-tiny.json +0 -160
- package/src/experimental/training/workload-packs/distill-glm52-doppler-code-agent-gemma270m-f16-sft-tiny.json +0 -175
- package/src/experimental/training/workload-packs/distill-glm52-doppler-code-agent-sft-tiny.json +0 -175
- package/src/experimental/training/workload-packs/distill-glm52-doppler-js-json-gemma270m-f16-sft-tiny.json +0 -160
- package/src/experimental/training/workload-packs/distill-glm52-doppler-review-gemma270m-f16-sft-tiny.json +0 -162
- package/src/experimental/training/workload-packs/distill-glm52-doppler-wgsl-gemma270m-f16-sft-tiny.json +0 -160
- package/src/experimental/training/workload-packs/distill-translategemma-medium.json +0 -103
- package/src/experimental/training/workload-packs/distill-translategemma-tiny.json +0 -103
- package/src/experimental/training/workload-packs/lora-doppler-agent-harness-gemma270m-f16-tiny.json +0 -137
- package/src/experimental/training/workload-packs/lora-doppler-code-agent-gemma270m-f16-tiny.json +0 -152
- package/src/experimental/training/workload-packs/lora-doppler-code-agent-tiny.json +0 -152
- package/src/experimental/training/workload-packs/lora-doppler-js-json-gemma270m-f16-tiny.json +0 -137
- package/src/experimental/training/workload-packs/lora-doppler-review-gemma270m-f16-tiny.json +0 -139
- package/src/experimental/training/workload-packs/lora-doppler-wgsl-gemma270m-f16-tiny.json +0 -137
- package/src/experimental/training/workload-packs/lora-toy-medium.json +0 -91
- package/src/experimental/training/workload-packs/lora-toy-tiny.json +0 -91
- package/src/experimental/training/workload-packs/registry.json +0 -278
- package/src/experimental/training/workload-packs/ul-training-medium.json +0 -55
- package/src/experimental/training/workload-packs/ul-training-tiny.json +0 -55
- package/src/experimental/training/workloads.d.ts +0 -201
- package/src/formats/gguf/types.d.ts +0 -74
- package/src/formats/index.d.ts +0 -75
- package/src/formats/index.js +0 -19
- package/src/formats/litert/types.d.ts +0 -74
- package/src/formats/rdrr/distributed/parsing.d.ts +0 -6
- package/src/formats/rdrr/distributed/validation.d.ts +0 -23
- package/src/formats/rdrr/storage-descriptor.d.ts +0 -36
- package/src/formats/rdrr/tensor-config-validator.d.ts +0 -33
- package/src/formats/safetensors/types.d.ts +0 -67
- package/src/formats/tflite/types.d.ts +0 -61
- package/src/formats/tokenizer/index.d.ts +0 -5
- package/src/formats/tokenizer/index.js +0 -3
- package/src/formats/tokenizer/types.d.ts +0 -9
- package/src/gpu/kernel-runtime.d.ts +0 -13
- package/src/gpu/kernel-selection-cache.d.ts +0 -13
- package/src/gpu/kernel-selection-log.d.ts +0 -4
- package/src/gpu/kernel-tuner/benchmarks.d.ts +0 -144
- package/src/gpu/kernel-tuner/cache.d.ts +0 -55
- package/src/gpu/kernel-tuner/index.d.ts +0 -59
- package/src/gpu/kernel-tuner/tuner.d.ts +0 -82
- package/src/gpu/kernel-tuner/types.d.ts +0 -101
- package/src/gpu/kernel-tuner/types.js +0 -4
- package/src/gpu/kernel-tuner.d.ts +0 -33
- package/src/gpu/kernels/README.md +0 -159
- package/src/gpu/kernels/check-finiteness.d.ts +0 -15
- package/src/gpu/kernels/check-hot-vocab-stop.d.ts +0 -18
- package/src/gpu/kernels/check-stop.d.ts +0 -31
- package/src/gpu/kernels/codegen/patches/attention_decode_f16.from.attention_decode.diff +0 -213
- package/src/gpu/kernels/codegen/patches/attention_decode_f16kv.from.attention_decode.diff +0 -89
- package/src/gpu/kernels/codegen/patches/attention_f16.from.attention.diff +0 -62
- package/src/gpu/kernels/codegen/patches/attention_f16kv.from.attention.diff +0 -40
- package/src/gpu/kernels/codegen/patches/attention_small_f16.from.attention_small.diff +0 -226
- package/src/gpu/kernels/codegen/patches/attention_small_f16kv.from.attention_small.diff +0 -264
- package/src/gpu/kernels/codegen/patches/attention_streaming_f16.from.attention_streaming.diff +0 -71
- package/src/gpu/kernels/codegen/patches/attention_streaming_f16kv.from.attention_streaming.diff +0 -53
- package/src/gpu/kernels/codegen/patches/bias_add_f16.from.bias_add.diff +0 -54
- package/src/gpu/kernels/codegen/patches/dequant_mxfp4_expert_f16.from.dequant_mxfp4_expert.diff +0 -125
- package/src/gpu/kernels/codegen/patches/fused_ffn_f16.from.fused_ffn.diff +0 -428
- package/src/gpu/kernels/codegen/patches/fused_matmul_q4_batched_f16.from.fused_matmul_q4_batched.diff +0 -114
- package/src/gpu/kernels/codegen/patches/fused_matmul_rmsnorm_f16.from.fused_matmul_rmsnorm.diff +0 -477
- package/src/gpu/kernels/codegen/patches/fused_swiglu_f16.from.fused_swiglu.diff +0 -60
- package/src/gpu/kernels/codegen/patches/gelu_f16.from.gelu.diff +0 -41
- package/src/gpu/kernels/codegen/patches/layernorm_f16.from.layernorm.diff +0 -82
- package/src/gpu/kernels/codegen/patches/matmul_gemv_residual_f16.from.matmul_gemv_residual.diff +0 -87
- package/src/gpu/kernels/codegen/patches/moe_gather_f16.from.moe_gather.diff +0 -181
- package/src/gpu/kernels/codegen/patches/residual_f16.from.residual.diff +0 -65
- package/src/gpu/kernels/codegen/patches/rmsnorm_f16.from.rmsnorm.diff +0 -515
- package/src/gpu/kernels/codegen/patches/rope_f16.from.rope.diff +0 -292
- package/src/gpu/kernels/codegen/patches/sample_f16.from.sample.diff +0 -336
- package/src/gpu/kernels/codegen/patches/scatter_add_dynamic_f16.from.scatter_add_dynamic.diff +0 -65
- package/src/gpu/kernels/codegen/patches/silu_f16.from.silu.diff +0 -81
- package/src/gpu/kernels/codegen/patches/topk_f16.from.topk.diff +0 -231
- package/src/gpu/kernels/codegen/wgsl-patch-variants.js +0 -152
- package/src/gpu/kernels/codegen/wgsl-variants.js +0 -885
- package/src/gpu/kernels/constants.d.ts +0 -129
- package/src/gpu/kernels/dispatch.d.ts +0 -72
- package/src/gpu/kernels/gated-short-conv.d.ts +0 -63
- package/src/gpu/kernels/kernel-base.d.ts +0 -33
- package/src/gpu/kernels/linear-attention-core.d.ts +0 -42
- package/src/gpu/kernels/matmul-dispatch.d.ts +0 -49
- package/src/gpu/kernels/matmul-selection.d.ts +0 -87
- package/src/gpu/kernels/rep-penalty.d.ts +0 -21
- package/src/gpu/kernels/rule-matcher.d.ts +0 -30
- package/src/gpu/kernels/rule-registry.d.ts +0 -7
- package/src/gpu/kernels/types.js +0 -4
- package/src/gpu/perf-guards.d.ts +0 -13
- package/src/gpu/profiler.d.ts +0 -93
- package/src/gpu/readback-utils.d.ts +0 -16
- package/src/gpu/readback-utils.js +0 -49
- package/src/gpu/submit-probe.d.ts +0 -12
- package/src/index-browser.d.ts +0 -2
- package/src/index-browser.js +0 -2
- package/src/inference/README.md +0 -593
- package/src/inference/browser-harness-contract-helpers.d.ts +0 -5
- package/src/inference/browser-harness-diffusion-energy-suites.d.ts +0 -2
- package/src/inference/browser-harness-model-helpers.d.ts +0 -10
- package/src/inference/browser-harness-report-helpers.d.ts +0 -7
- package/src/inference/browser-harness-suite-helpers.d.ts +0 -40
- package/src/inference/browser-harness-text-helpers.d.ts +0 -124
- package/src/inference/model-lanes/execution-lane-resolver.d.ts +0 -61
- package/src/inference/model-lanes/execution-lane-resolver.js +0 -126
- package/src/inference/network-evolution.js +0 -79
- package/src/inference/pipelines/context.d.ts +0 -21
- package/src/inference/pipelines/diffusion/helpers.d.ts +0 -29
- package/src/inference/pipelines/diffusion/index.d.ts +0 -3
- package/src/inference/pipelines/diffusion/index.js +0 -3
- package/src/inference/pipelines/diffusion/scheduler.d.ts +0 -35
- package/src/inference/pipelines/diffusion/sd3-transformer.d.ts +0 -20
- package/src/inference/pipelines/diffusion/sd3-weights.d.ts +0 -21
- package/src/inference/pipelines/diffusion/text-encoder-gpu.d.ts +0 -82
- package/src/inference/pipelines/diffusion/text-encoder.d.ts +0 -29
- package/src/inference/pipelines/diffusion/types.js +0 -1
- package/src/inference/pipelines/diffusion/vae.d.ts +0 -20
- package/src/inference/pipelines/dream/energy-head-pipeline.d.ts +0 -29
- package/src/inference/pipelines/dream/energy-head-pipeline.js +0 -6
- package/src/inference/pipelines/dream/pipeline.d.ts +0 -17
- package/src/inference/pipelines/dream/pipeline.js +0 -8
- package/src/inference/pipelines/energy/index.d.ts +0 -1
- package/src/inference/pipelines/energy/index.js +0 -1
- package/src/inference/pipelines/energy/types.js +0 -1
- package/src/inference/pipelines/energy-head/index.d.ts +0 -6
- package/src/inference/pipelines/energy-head/index.js +0 -6
- package/src/inference/pipelines/factory.d.ts +0 -10
- package/src/inference/pipelines/index.d.ts +0 -22
- package/src/inference/pipelines/index.js +0 -19
- package/src/inference/pipelines/registry.d.ts +0 -15
- package/src/inference/pipelines/rng.d.ts +0 -2
- package/src/inference/pipelines/structured/tsv-grammar-mask.js +0 -166
- package/src/inference/pipelines/text/attention/attn-config.d.ts +0 -23
- package/src/inference/pipelines/text/attention/dispatch-params.d.ts +0 -173
- package/src/inference/pipelines/text/attention/index.d.ts +0 -24
- package/src/inference/pipelines/text/attention/output-projection.d.ts +0 -12
- package/src/inference/pipelines/text/attention/precision-contract.d.ts +0 -36
- package/src/inference/pipelines/text/attention/projections.d.ts +0 -208
- package/src/inference/pipelines/text/attention/quant-options.d.ts +0 -38
- package/src/inference/pipelines/text/attention/record.d.ts +0 -36
- package/src/inference/pipelines/text/attention/rmsnorm-fusion-gate.d.ts +0 -8
- package/src/inference/pipelines/text/attention/run.d.ts +0 -38
- package/src/inference/pipelines/text/attention/types.d.ts +0 -125
- package/src/inference/pipelines/text/attention.d.ts +0 -23
- package/src/inference/pipelines/text/attention.js +0 -12
- package/src/inference/pipelines/text/bdpa-steamroller.d.ts +0 -22
- package/src/inference/pipelines/text/buffer-types.js +0 -4
- package/src/inference/pipelines/text/debug-utils/config.d.ts +0 -144
- package/src/inference/pipelines/text/debug-utils/index.d.ts +0 -53
- package/src/inference/pipelines/text/debug-utils/logging.d.ts +0 -106
- package/src/inference/pipelines/text/debug-utils/tensor.d.ts +0 -119
- package/src/inference/pipelines/text/debug-utils/utils.d.ts +0 -77
- package/src/inference/pipelines/text/debug-utils.d.ts +0 -42
- package/src/inference/pipelines/text/drift-policy.d.ts +0 -57
- package/src/inference/pipelines/text/dtype-contract.d.ts +0 -53
- package/src/inference/pipelines/text/embed.d.ts +0 -68
- package/src/inference/pipelines/text/execution-runtime-builders.d.ts +0 -49
- package/src/inference/pipelines/text/failure-reduction.d.ts +0 -84
- package/src/inference/pipelines/text/failure-reduction.js +0 -180
- package/src/inference/pipelines/text/ffn/dense.d.ts +0 -99
- package/src/inference/pipelines/text/ffn/index.d.ts +0 -23
- package/src/inference/pipelines/text/ffn/moe.d.ts +0 -24
- package/src/inference/pipelines/text/ffn/sandwich.d.ts +0 -27
- package/src/inference/pipelines/text/ffn/standard.d.ts +0 -25
- package/src/inference/pipelines/text/ffn/types.d.ts +0 -30
- package/src/inference/pipelines/text/ffn.d.ts +0 -31
- package/src/inference/pipelines/text/ffn.js +0 -18
- package/src/inference/pipelines/text/finiteness-guard-status.d.ts +0 -11
- package/src/inference/pipelines/text/finiteness-policy.d.ts +0 -39
- package/src/inference/pipelines/text/generator-decode-policy.d.ts +0 -81
- package/src/inference/pipelines/text/generator-helpers.d.ts +0 -45
- package/src/inference/pipelines/text/generator-prefill-helpers.d.ts +0 -84
- package/src/inference/pipelines/text/generator-runtime.d.ts +0 -89
- package/src/inference/pipelines/text/generator-steps.d.ts +0 -186
- package/src/inference/pipelines/text/index.d.ts +0 -5
- package/src/inference/pipelines/text/index.js +0 -6
- package/src/inference/pipelines/text/kernel-trace.d.ts +0 -119
- package/src/inference/pipelines/text/layer.d.ts +0 -118
- package/src/inference/pipelines/text/load-errors.d.ts +0 -21
- package/src/inference/pipelines/text/logits/types.js +0 -4
- package/src/inference/pipelines/text/logits.d.ts +0 -27
- package/src/inference/pipelines/text/logits.js +0 -16
- package/src/inference/pipelines/text/lora-apply.d.ts +0 -28
- package/src/inference/pipelines/text/moe-cache.d.ts +0 -32
- package/src/inference/pipelines/text/moe-cpu-gptoss.d.ts +0 -9
- package/src/inference/pipelines/text/moe-cpu.d.ts +0 -13
- package/src/inference/pipelines/text/moe-gpu.d.ts +0 -26
- package/src/inference/pipelines/text/moe-helpers.d.ts +0 -12
- package/src/inference/pipelines/text/moe-shape-validator.d.ts +0 -40
- package/src/inference/pipelines/text/operator-events.d.ts +0 -125
- package/src/inference/pipelines/text/operator-identity.d.ts +0 -60
- package/src/inference/pipelines/text/ops.d.ts +0 -219
- package/src/inference/pipelines/text/per-layer-inputs.d.ts +0 -107
- package/src/inference/pipelines/text/probes.d.ts +0 -28
- package/src/inference/pipelines/text/sampling-config.d.ts +0 -48
- package/src/inference/pipelines/text/stage-names.d.ts +0 -88
- package/src/inference/pipelines/text/tsir-fixture-writer.d.ts +0 -57
- package/src/inference/pipelines/text/types.js +0 -4
- package/src/inference/tokenizers/base.d.ts +0 -44
- package/src/inference/tokenizers/behavior-flags.d.ts +0 -9
- package/src/inference/tokenizers/bpe.d.ts +0 -27
- package/src/inference/tokenizers/bundled.d.ts +0 -67
- package/src/inference/tokenizers/sentencepiece.d.ts +0 -30
- package/src/inference/tokenizers/types.js +0 -7
- package/src/loader/embedding-limit-preflight.d.ts +0 -31
- package/src/loader/per-layer-input-loader.d.ts +0 -30
- package/src/loader/tensors/functional-descriptor-loader.d.ts +0 -9
- package/src/loader/tensors/source-transform.d.ts +0 -65
- package/src/memory/address-table.d.ts +0 -96
- package/src/rules/execution-rules-contract-check.d.ts +0 -21
- package/src/rules/layer-pattern-contract-check.d.ts +0 -17
- package/src/rules/rule-registry.d.ts +0 -77
- package/src/storage/artifact-storage-context.d.ts +0 -63
- package/src/storage/backends/idb-store.d.ts +0 -52
- package/src/storage/backends/memory-store.d.ts +0 -36
- package/src/storage/backends/opfs-store.d.ts +0 -41
- package/src/storage/blake3.d.ts +0 -17
- package/src/storage/download-types.d.ts +0 -157
- package/src/storage/downloader.d.ts +0 -103
- package/src/storage/index.d.ts +0 -274
- package/src/storage/index.js +0 -201
- package/src/storage/preflight.d.ts +0 -144
- package/src/storage/preflight.js +0 -360
- package/src/storage/quickstart-downloader.d.ts +0 -157
- package/src/storage/quickstart-downloader.js +0 -289
- package/src/storage/source-artifact-store.d.ts +0 -59
- package/src/sw.js +0 -184
- package/src/tooling/command-api-constants.d.ts +0 -9
- package/src/tooling/command-api-family-normalizers.d.ts +0 -10
- package/src/tooling/command-api-helpers.d.ts +0 -23
- package/src/tooling/command-runner-shared.d.ts +0 -89
- package/src/tooling/conversion-config-materializer.d.ts +0 -14
- package/src/tooling/conversion-config-materializer.js +0 -96
- package/src/tooling/diagnose-runner.d.ts +0 -28
- package/src/tooling/hf-registry-utils.d.ts +0 -55
- package/src/tooling/kernel-path-builder/custom-runtime-facts.d.ts +0 -18
- package/src/tooling/kernel-path-builder/custom-runtime-facts.js +0 -72
- package/src/tooling/kernel-path-builder/index.d.ts +0 -65
- package/src/tooling/kernel-path-builder/index.js +0 -1361
- package/src/tooling/kernel-path-builder/runtime-overlay.d.ts +0 -9
- package/src/tooling/kernel-path-builder/runtime-overlay.js +0 -147
- package/src/tooling/lean-execution-contract-runner.d.ts +0 -33
- package/src/tooling/lean-execution-contract-runner.js +0 -158
- package/src/tooling/lean-execution-contract.d.ts +0 -16
- package/src/tooling/lean-execution-contract.js +0 -81
- package/src/tooling/litert-package-runtime.d.ts +0 -46
- package/src/tooling/node-convert-worker-pool.d.ts +0 -16
- package/src/tooling/node-convert-worker.d.ts +0 -1
- package/src/tooling/node-converter.d.ts +0 -1
- package/src/tooling/node-file-fetch.d.ts +0 -1
- package/src/tooling/node-source-runtime.d.ts +0 -24
- package/src/tooling/node-webgpu.d.ts +0 -17
- package/src/tooling/precision-replay-math.js +0 -174
- package/src/tooling/program-bundle-host.d.ts +0 -16
- package/src/tooling/program-bundle-host.js +0 -20
- package/src/tooling/runtime-input-composition.d.ts +0 -37
- package/src/tooling/source-package-profiles.d.ts +0 -50
- package/src/tooling/source-runtime-converter-config.d.ts +0 -19
- package/src/tooling/source-runtime-materializer.d.ts +0 -6
- package/src/tooling/source-runtime-materializer.js +0 -89
- package/src/tooling-exports.browser.d.ts +0 -7
- package/src/types/chrome.d.ts +0 -36
- package/src/types/chrome.js +0 -1
- package/src/types/index.d.ts +0 -1
- package/src/types/index.js +0 -1
- package/src/utils/array-buffer.d.ts +0 -4
- package/src/utils/canonical-hash.d.ts +0 -7
- package/src/utils/clone-json.d.ts +0 -1
- package/src/utils/hf-resolve-url.d.ts +0 -16
- package/src/utils/index.d.ts +0 -10
- package/src/utils/index.js +0 -10
- package/src/utils/load-json.d.ts +0 -5
- package/src/utils/plain-object.d.ts +0 -1
- package/src/utils/sha256.d.ts +0 -5
- package/src/utils/stable-sort-object.d.ts +0 -1
|
@@ -1,268 +0,0 @@
|
|
|
1
|
-
#!/usr/bin/env python3
|
|
2
|
-
"""
|
|
3
|
-
Dump intermediate values from Qwen3.5 linear attention (GatedDeltaNet) for comparison with Doppler.
|
|
4
|
-
|
|
5
|
-
Usage:
|
|
6
|
-
HF_HOME=/media/x/models/huggingface_cache python3 src/debug/reference/hf_qwen35_linear_attn_debug.py
|
|
7
|
-
"""
|
|
8
|
-
|
|
9
|
-
import os
|
|
10
|
-
import torch
|
|
11
|
-
import numpy as np
|
|
12
|
-
|
|
13
|
-
os.environ.setdefault("HF_HOME", "/media/x/models/huggingface_cache")
|
|
14
|
-
|
|
15
|
-
from transformers import AutoModelForCausalLM, AutoTokenizer
|
|
16
|
-
|
|
17
|
-
MODEL_ID = "Qwen/Qwen3.5-0.8B"
|
|
18
|
-
PROMPT = "Hello"
|
|
19
|
-
|
|
20
|
-
|
|
21
|
-
def stats(name, tensor):
|
|
22
|
-
t = tensor.float().detach().flatten()
|
|
23
|
-
print(f" {name}: shape={list(tensor.shape)}, "
|
|
24
|
-
f"min={t.min().item():.6f}, max={t.max().item():.6f}, "
|
|
25
|
-
f"mean={t.mean().item():.6f}, absMax={t.abs().max().item():.6f}")
|
|
26
|
-
first8 = t[:8].tolist()
|
|
27
|
-
print(f" first8: {[f'{v:.6f}' for v in first8]}")
|
|
28
|
-
|
|
29
|
-
|
|
30
|
-
def main():
|
|
31
|
-
print(f"Loading {MODEL_ID}...")
|
|
32
|
-
model = AutoModelForCausalLM.from_pretrained(MODEL_ID, dtype=torch.float32)
|
|
33
|
-
tokenizer = AutoTokenizer.from_pretrained(MODEL_ID)
|
|
34
|
-
model.eval()
|
|
35
|
-
|
|
36
|
-
inputs = tokenizer(PROMPT, return_tensors="pt")
|
|
37
|
-
input_ids = inputs["input_ids"]
|
|
38
|
-
print(f"Prompt: '{PROMPT}', Token IDs: {input_ids[0].tolist()}")
|
|
39
|
-
num_tokens = input_ids.shape[1]
|
|
40
|
-
|
|
41
|
-
# Dump key weight values for layer 0
|
|
42
|
-
layer0 = model.model.layers[0]
|
|
43
|
-
attn = layer0.linear_attn
|
|
44
|
-
|
|
45
|
-
print(f"\n=== Layer 0 weights ===")
|
|
46
|
-
if hasattr(attn, 'A_log'):
|
|
47
|
-
a_log = attn.A_log.detach().float()
|
|
48
|
-
a_neg_exp = -torch.exp(a_log)
|
|
49
|
-
stats("A_log", a_log)
|
|
50
|
-
stats("a_neg_exp", a_neg_exp)
|
|
51
|
-
if hasattr(attn, 'dt_bias'):
|
|
52
|
-
stats("dt_bias", attn.dt_bias.detach().float())
|
|
53
|
-
stats("conv1d.weight", attn.conv1d.weight.detach().float())
|
|
54
|
-
stats("norm.weight", attn.norm.weight.detach().float())
|
|
55
|
-
|
|
56
|
-
# Hook into the linear_attn module to capture its input and output
|
|
57
|
-
captured = {}
|
|
58
|
-
|
|
59
|
-
def hook_linear_attn_input(module, args, kwargs):
|
|
60
|
-
if len(args) > 0:
|
|
61
|
-
captured['linear_attn_input'] = args[0].detach().clone()
|
|
62
|
-
return None
|
|
63
|
-
|
|
64
|
-
def hook_linear_attn_output(module, args, kwargs, output):
|
|
65
|
-
if isinstance(output, tuple):
|
|
66
|
-
captured['linear_attn_output'] = output[0].detach().clone()
|
|
67
|
-
else:
|
|
68
|
-
captured['linear_attn_output'] = output.detach().clone()
|
|
69
|
-
return None
|
|
70
|
-
|
|
71
|
-
# Hook into individual projection layers
|
|
72
|
-
def make_hook(name):
|
|
73
|
-
def hook(module, input, output):
|
|
74
|
-
captured[name] = output.detach().clone()
|
|
75
|
-
return hook
|
|
76
|
-
|
|
77
|
-
hooks = []
|
|
78
|
-
hooks.append(attn.register_forward_pre_hook(hook_linear_attn_input, with_kwargs=True))
|
|
79
|
-
hooks.append(attn.register_forward_hook(hook_linear_attn_output, with_kwargs=True))
|
|
80
|
-
hooks.append(attn.in_proj_qkv.register_forward_hook(make_hook('qkv_proj')))
|
|
81
|
-
hooks.append(attn.in_proj_z.register_forward_hook(make_hook('z_proj')))
|
|
82
|
-
hooks.append(attn.in_proj_a.register_forward_hook(make_hook('a_proj')))
|
|
83
|
-
hooks.append(attn.in_proj_b.register_forward_hook(make_hook('b_proj')))
|
|
84
|
-
hooks.append(attn.out_proj.register_forward_hook(make_hook('out_proj')))
|
|
85
|
-
hooks.append(attn.conv1d.register_forward_hook(make_hook('conv1d_raw')))
|
|
86
|
-
hooks.append(attn.norm.register_forward_hook(make_hook('gated_norm')))
|
|
87
|
-
|
|
88
|
-
# Also hook input_layernorm
|
|
89
|
-
hooks.append(layer0.input_layernorm.register_forward_hook(make_hook('input_layernorm')))
|
|
90
|
-
|
|
91
|
-
print(f"\n=== Running forward pass ===")
|
|
92
|
-
with torch.no_grad():
|
|
93
|
-
outputs = model(input_ids, output_hidden_states=True)
|
|
94
|
-
|
|
95
|
-
# Remove hooks
|
|
96
|
-
for h in hooks:
|
|
97
|
-
h.remove()
|
|
98
|
-
|
|
99
|
-
print(f"\n=== Captured intermediates ===")
|
|
100
|
-
for name in ['input_layernorm', 'qkv_proj', 'z_proj', 'a_proj', 'b_proj',
|
|
101
|
-
'conv1d_raw', 'gated_norm', 'linear_attn_input', 'linear_attn_output', 'out_proj']:
|
|
102
|
-
if name in captured:
|
|
103
|
-
stats(name, captured[name])
|
|
104
|
-
else:
|
|
105
|
-
print(f" {name}: NOT CAPTURED")
|
|
106
|
-
|
|
107
|
-
# Hidden states per layer
|
|
108
|
-
print(f"\n=== Hidden states per layer (last token) ===")
|
|
109
|
-
for i in range(min(6, len(outputs.hidden_states) - 1)):
|
|
110
|
-
hs = outputs.hidden_states[i + 1]
|
|
111
|
-
t = hs[0, -1] # last token
|
|
112
|
-
vals = t[:8].tolist()
|
|
113
|
-
max_abs = t.abs().max().item()
|
|
114
|
-
mean_abs = t.abs().mean().item()
|
|
115
|
-
layer_type = type(model.model.layers[i]).__name__
|
|
116
|
-
attn_type = "linear" if hasattr(model.model.layers[i], 'linear_attn') else "full"
|
|
117
|
-
print(f" Layer {i} ({attn_type}): first8={[f'{v:.4f}' for v in vals]}, "
|
|
118
|
-
f"maxAbs={max_abs:.4f}, meanAbs={mean_abs:.4f}")
|
|
119
|
-
|
|
120
|
-
# Logits
|
|
121
|
-
logits = outputs.logits[0, -1]
|
|
122
|
-
top5 = torch.topk(logits, 5)
|
|
123
|
-
print(f"\nTop-5 logits: {[(tokenizer.decode([idx.item()]), f'{val.item():.2f}') for val, idx in zip(top5.values, top5.indices)]}")
|
|
124
|
-
|
|
125
|
-
# Also trace through the linear attention manually to compare with Doppler's kernel
|
|
126
|
-
print(f"\n=== Manual linear attention trace (layer 0) ===")
|
|
127
|
-
with torch.no_grad():
|
|
128
|
-
embed = model.model.embed_tokens(input_ids)
|
|
129
|
-
normed = layer0.input_layernorm(embed)
|
|
130
|
-
stats("normed_input", normed)
|
|
131
|
-
|
|
132
|
-
qkv = attn.in_proj_qkv(normed)
|
|
133
|
-
stats("qkv", qkv)
|
|
134
|
-
|
|
135
|
-
# The HF Qwen3.5 GatedDeltaNet does conv1d on the QKV, then applies SiLU
|
|
136
|
-
# The conv1d expects [batch, channels, seq_len] format
|
|
137
|
-
qkv_t = qkv.transpose(1, 2) # [1, 6144, 1]
|
|
138
|
-
|
|
139
|
-
# Use the conv1d module directly (it has padding configured)
|
|
140
|
-
conv_raw = attn.conv1d(qkv_t)
|
|
141
|
-
stats("conv_raw (from module)", conv_raw.transpose(1, 2))
|
|
142
|
-
|
|
143
|
-
# Truncate to seq_len (causal conv padding)
|
|
144
|
-
conv_causal = conv_raw[..., :num_tokens]
|
|
145
|
-
stats("conv_causal (truncated)", conv_causal.transpose(1, 2))
|
|
146
|
-
|
|
147
|
-
# Apply SiLU
|
|
148
|
-
conv_silu = torch.nn.functional.silu(conv_causal)
|
|
149
|
-
stats("conv_silu", conv_silu.transpose(1, 2))
|
|
150
|
-
|
|
151
|
-
# Split Q, K, V
|
|
152
|
-
conv_out = conv_silu.transpose(1, 2) # [1, seq_len, 6144]
|
|
153
|
-
num_k_heads = 16
|
|
154
|
-
head_k_dim = 128
|
|
155
|
-
head_v_dim = 128
|
|
156
|
-
num_v_heads = 16
|
|
157
|
-
q_size = num_k_heads * head_k_dim # 2048
|
|
158
|
-
k_size = q_size
|
|
159
|
-
v_size = num_v_heads * head_v_dim # 2048
|
|
160
|
-
|
|
161
|
-
q = conv_out[..., :q_size]
|
|
162
|
-
k = conv_out[..., q_size:q_size + k_size]
|
|
163
|
-
v = conv_out[..., q_size + k_size:]
|
|
164
|
-
stats("Q (raw)", q)
|
|
165
|
-
stats("K (raw)", k)
|
|
166
|
-
stats("V (raw)", v)
|
|
167
|
-
|
|
168
|
-
# Reshape for per-head processing
|
|
169
|
-
# Q and K: [batch, seq, num_k_heads, head_k_dim]
|
|
170
|
-
q_heads = q.view(1, num_tokens, num_k_heads, head_k_dim)
|
|
171
|
-
k_heads = k.view(1, num_tokens, num_k_heads, head_k_dim)
|
|
172
|
-
v_heads = v.view(1, num_tokens, num_v_heads, head_v_dim)
|
|
173
|
-
|
|
174
|
-
# L2 normalize Q and K
|
|
175
|
-
eps = 1e-6
|
|
176
|
-
q_norm = torch.nn.functional.normalize(q_heads, p=2, dim=-1, eps=eps)
|
|
177
|
-
k_norm = torch.nn.functional.normalize(k_heads, p=2, dim=-1, eps=eps)
|
|
178
|
-
|
|
179
|
-
# Scale Q by 1/sqrt(head_k_dim)
|
|
180
|
-
head_scale = 1.0 / (head_k_dim ** 0.5)
|
|
181
|
-
q_scaled = q_norm * head_scale
|
|
182
|
-
|
|
183
|
-
stats("Q_normed_scaled (per-head)", q_scaled.reshape(1, num_tokens, -1))
|
|
184
|
-
stats("K_normed (per-head)", k_norm.reshape(1, num_tokens, -1))
|
|
185
|
-
|
|
186
|
-
# Projections for gating
|
|
187
|
-
z = attn.in_proj_z(normed)
|
|
188
|
-
a_out = attn.in_proj_a(normed)
|
|
189
|
-
b_out = attn.in_proj_b(normed)
|
|
190
|
-
stats("z", z)
|
|
191
|
-
stats("a", a_out)
|
|
192
|
-
stats("b", b_out)
|
|
193
|
-
|
|
194
|
-
# Compute gating values
|
|
195
|
-
a_log = attn.A_log.detach().float()
|
|
196
|
-
a_neg_exp = -torch.exp(a_log)
|
|
197
|
-
dt_bias = attn.dt_bias.detach().float()
|
|
198
|
-
|
|
199
|
-
softplus_input = a_out.squeeze(0).squeeze(0) + dt_bias
|
|
200
|
-
softplus_val = torch.nn.functional.softplus(softplus_input)
|
|
201
|
-
g = a_neg_exp * softplus_val
|
|
202
|
-
g_exp = torch.exp(g)
|
|
203
|
-
beta = torch.sigmoid(b_out.squeeze(0).squeeze(0))
|
|
204
|
-
|
|
205
|
-
stats("softplus(a + dt_bias)", softplus_val.unsqueeze(0).unsqueeze(0))
|
|
206
|
-
stats("g (decay)", g.unsqueeze(0).unsqueeze(0))
|
|
207
|
-
stats("g_exp (decay factor)", g_exp.unsqueeze(0).unsqueeze(0))
|
|
208
|
-
stats("beta (sigmoid(b))", beta.unsqueeze(0).unsqueeze(0))
|
|
209
|
-
|
|
210
|
-
# Recurrent state update (for first token, state is all zeros)
|
|
211
|
-
# state[head, kd, vd] = state * g_exp + k[kd] * delta[vd]
|
|
212
|
-
# where delta[vd] = (v[vd] - state^T @ k * beta
|
|
213
|
-
# For zero state: delta[vd] = v[vd] * beta, state = k ⊗ delta
|
|
214
|
-
state = torch.zeros(num_v_heads, head_k_dim, head_v_dim)
|
|
215
|
-
|
|
216
|
-
# Apply decay (no-op for zero state)
|
|
217
|
-
for head in range(num_v_heads):
|
|
218
|
-
state[head] *= g_exp[head].item()
|
|
219
|
-
|
|
220
|
-
k_head = k_norm[0, 0, head % num_k_heads] # broadcast q_rep
|
|
221
|
-
v_head = v_heads[0, 0, head]
|
|
222
|
-
|
|
223
|
-
# kv_mem = state @ k
|
|
224
|
-
kv_mem = state[head].t() @ k_head # [head_v_dim]
|
|
225
|
-
|
|
226
|
-
# delta = (v - kv_mem) * beta
|
|
227
|
-
delta = (v_head - kv_mem) * beta[head].item()
|
|
228
|
-
|
|
229
|
-
# state += outer(k, delta)
|
|
230
|
-
state[head] += torch.outer(k_head, delta)
|
|
231
|
-
|
|
232
|
-
# Output: out = state^T @ q
|
|
233
|
-
output_per_head = torch.zeros(1, num_tokens, num_v_heads, head_v_dim)
|
|
234
|
-
for head in range(num_v_heads):
|
|
235
|
-
q_head = q_scaled[0, 0, head % num_k_heads]
|
|
236
|
-
out_head = state[head].t() @ q_head # [head_v_dim]
|
|
237
|
-
output_per_head[0, 0, head] = out_head
|
|
238
|
-
|
|
239
|
-
raw_out = output_per_head.reshape(1, num_tokens, num_v_heads * head_v_dim)
|
|
240
|
-
stats("Recurrent output (raw)", raw_out)
|
|
241
|
-
|
|
242
|
-
# RMS norm per head + SiLU gate
|
|
243
|
-
z_reshaped = z.view(1, num_tokens, num_v_heads, head_v_dim)
|
|
244
|
-
norm_weight = attn.norm.weight.detach().float() # [head_v_dim] (shared mode)
|
|
245
|
-
rms_eps = 1e-6
|
|
246
|
-
|
|
247
|
-
for head in range(num_v_heads):
|
|
248
|
-
head_out = output_per_head[0, 0, head] # [head_v_dim]
|
|
249
|
-
mean_sq = (head_out ** 2).mean()
|
|
250
|
-
inv_rms = 1.0 / torch.sqrt(mean_sq + rms_eps)
|
|
251
|
-
z_gate = torch.nn.functional.silu(z_reshaped[0, 0, head])
|
|
252
|
-
output_per_head[0, 0, head] = head_out * inv_rms * norm_weight * z_gate
|
|
253
|
-
|
|
254
|
-
gated_out = output_per_head.reshape(1, num_tokens, num_v_heads * head_v_dim)
|
|
255
|
-
stats("After RMSNorm + SiLU gate", gated_out)
|
|
256
|
-
|
|
257
|
-
# Output projection
|
|
258
|
-
o_result = torch.nn.functional.linear(gated_out, attn.out_proj.weight)
|
|
259
|
-
stats("After out_proj", o_result)
|
|
260
|
-
|
|
261
|
-
# Compare with captured output
|
|
262
|
-
if 'linear_attn_output' in captured:
|
|
263
|
-
diff = (o_result - captured['linear_attn_output']).abs()
|
|
264
|
-
print(f"\n Diff vs captured output: maxDiff={diff.max().item():.6f}")
|
|
265
|
-
|
|
266
|
-
|
|
267
|
-
if __name__ == "__main__":
|
|
268
|
-
main()
|
|
@@ -1,116 +0,0 @@
|
|
|
1
|
-
#!/usr/bin/env python3
|
|
2
|
-
"""
|
|
3
|
-
Verify RoPE (Rotary Position Embedding) values match between HF and DOPPLER.
|
|
4
|
-
|
|
5
|
-
Usage:
|
|
6
|
-
python hf_rope_check.py <config.json>
|
|
7
|
-
|
|
8
|
-
Config (JSON):
|
|
9
|
-
{
|
|
10
|
-
"model": "google/gemma-2-2b-it",
|
|
11
|
-
"pos": 6,
|
|
12
|
-
"dim": 256,
|
|
13
|
-
"theta": 10000.0
|
|
14
|
-
}
|
|
15
|
-
|
|
16
|
-
Requires: pip install torch transformers
|
|
17
|
-
"""
|
|
18
|
-
|
|
19
|
-
import json
|
|
20
|
-
import math
|
|
21
|
-
import sys
|
|
22
|
-
import torch
|
|
23
|
-
from transformers import AutoModelForCausalLM
|
|
24
|
-
|
|
25
|
-
|
|
26
|
-
def compute_rope_freqs(head_dim: int, theta: float = 10000.0):
|
|
27
|
-
"""Compute RoPE frequencies for a given head dimension."""
|
|
28
|
-
freqs = 1.0 / (theta ** (torch.arange(0, head_dim, 2).float() / head_dim))
|
|
29
|
-
return freqs
|
|
30
|
-
|
|
31
|
-
|
|
32
|
-
def apply_rope_to_vector(x: torch.Tensor, pos: int, freqs: torch.Tensor):
|
|
33
|
-
"""Apply RoPE to a single vector at a given position."""
|
|
34
|
-
# x shape: [head_dim]
|
|
35
|
-
angles = pos * freqs
|
|
36
|
-
cos = torch.cos(angles)
|
|
37
|
-
sin = torch.sin(angles)
|
|
38
|
-
|
|
39
|
-
# Split into pairs
|
|
40
|
-
x_even = x[0::2]
|
|
41
|
-
x_odd = x[1::2]
|
|
42
|
-
|
|
43
|
-
# Apply rotation
|
|
44
|
-
out_even = x_even * cos - x_odd * sin
|
|
45
|
-
out_odd = x_even * sin + x_odd * cos
|
|
46
|
-
|
|
47
|
-
# Interleave back
|
|
48
|
-
out = torch.zeros_like(x)
|
|
49
|
-
out[0::2] = out_even
|
|
50
|
-
out[1::2] = out_odd
|
|
51
|
-
|
|
52
|
-
return out
|
|
53
|
-
|
|
54
|
-
|
|
55
|
-
def main():
|
|
56
|
-
if len(sys.argv) != 2:
|
|
57
|
-
raise SystemExit("Usage: python hf_rope_check.py <config.json>")
|
|
58
|
-
path = sys.argv[1]
|
|
59
|
-
with open(path, "r", encoding="utf-8") as handle:
|
|
60
|
-
config = json.load(handle)
|
|
61
|
-
if not isinstance(config, dict):
|
|
62
|
-
raise SystemExit("Config must be a JSON object")
|
|
63
|
-
|
|
64
|
-
model_id = config.get("model")
|
|
65
|
-
pos = config.get("pos")
|
|
66
|
-
dim = config.get("dim")
|
|
67
|
-
theta = config.get("theta")
|
|
68
|
-
|
|
69
|
-
if not isinstance(model_id, str) or not model_id.strip():
|
|
70
|
-
raise SystemExit('Config "model" must be a non-empty string')
|
|
71
|
-
if not isinstance(pos, int):
|
|
72
|
-
raise SystemExit('Config "pos" must be an integer')
|
|
73
|
-
if not isinstance(dim, int):
|
|
74
|
-
raise SystemExit('Config "dim" must be an integer')
|
|
75
|
-
if not isinstance(theta, (int, float)):
|
|
76
|
-
raise SystemExit('Config "theta" must be a number')
|
|
77
|
-
|
|
78
|
-
print(f"RoPE Check for position {pos}, head_dim={dim}, theta={theta}")
|
|
79
|
-
|
|
80
|
-
# Compute frequencies
|
|
81
|
-
freqs = compute_rope_freqs(dim, theta)
|
|
82
|
-
print(f"\nFrequencies (first 8): {freqs[:8].tolist()}")
|
|
83
|
-
|
|
84
|
-
# Compute angles at position
|
|
85
|
-
angles = pos * freqs
|
|
86
|
-
print(f"Angles at pos {pos} (first 8): {angles[:8].tolist()}")
|
|
87
|
-
|
|
88
|
-
# Compute cos/sin
|
|
89
|
-
cos = torch.cos(angles)
|
|
90
|
-
sin = torch.sin(angles)
|
|
91
|
-
print(f"Cos at pos {pos} (first 8): {cos[:8].tolist()}")
|
|
92
|
-
print(f"Sin at pos {pos} (first 8): {sin[:8].tolist()}")
|
|
93
|
-
|
|
94
|
-
# Test with a sample vector
|
|
95
|
-
test_vec = torch.randn(dim)
|
|
96
|
-
test_vec[0] = 1.0
|
|
97
|
-
test_vec[1] = 0.0
|
|
98
|
-
print(f"\nTest vector (first 8): {test_vec[:8].tolist()}")
|
|
99
|
-
|
|
100
|
-
rotated = apply_rope_to_vector(test_vec, pos, freqs)
|
|
101
|
-
print(f"After RoPE (first 8): {rotated[:8].tolist()}")
|
|
102
|
-
|
|
103
|
-
# Also load model and check actual RoPE config
|
|
104
|
-
print(f"\nLoading model to verify config: {model_id}")
|
|
105
|
-
model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype=torch.float32, device_map="cpu")
|
|
106
|
-
config = model.config
|
|
107
|
-
|
|
108
|
-
print(f"\nModel RoPE config:")
|
|
109
|
-
print(f" rope_theta: {getattr(config, 'rope_theta', 'not set')}")
|
|
110
|
-
print(f" head_dim: {config.head_dim}")
|
|
111
|
-
if hasattr(config, 'rope_scaling'):
|
|
112
|
-
print(f" rope_scaling: {config.rope_scaling}")
|
|
113
|
-
|
|
114
|
-
|
|
115
|
-
if __name__ == "__main__":
|
|
116
|
-
main()
|
|
@@ -1,75 +0,0 @@
|
|
|
1
|
-
#!/usr/bin/env python3
|
|
2
|
-
"""
|
|
3
|
-
Dump weight values for Q/K/V/O projections from HuggingFace model.
|
|
4
|
-
|
|
5
|
-
Usage:
|
|
6
|
-
python hf_weights.py <config.json>
|
|
7
|
-
|
|
8
|
-
Config (JSON):
|
|
9
|
-
{
|
|
10
|
-
"model": "google/gemma-2-2b-it",
|
|
11
|
-
"layer": 0,
|
|
12
|
-
"proj": "all"
|
|
13
|
-
}
|
|
14
|
-
|
|
15
|
-
Requires: pip install torch transformers
|
|
16
|
-
"""
|
|
17
|
-
|
|
18
|
-
import json
|
|
19
|
-
import sys
|
|
20
|
-
import torch
|
|
21
|
-
from transformers import AutoModelForCausalLM
|
|
22
|
-
|
|
23
|
-
|
|
24
|
-
def main():
|
|
25
|
-
if len(sys.argv) != 2:
|
|
26
|
-
raise SystemExit("Usage: python hf_weights.py <config.json>")
|
|
27
|
-
path = sys.argv[1]
|
|
28
|
-
with open(path, "r", encoding="utf-8") as handle:
|
|
29
|
-
config = json.load(handle)
|
|
30
|
-
if not isinstance(config, dict):
|
|
31
|
-
raise SystemExit("Config must be a JSON object")
|
|
32
|
-
|
|
33
|
-
model_id = config.get("model")
|
|
34
|
-
layer_index = config.get("layer")
|
|
35
|
-
proj = config.get("proj")
|
|
36
|
-
|
|
37
|
-
if not isinstance(model_id, str) or not model_id.strip():
|
|
38
|
-
raise SystemExit('Config "model" must be a non-empty string')
|
|
39
|
-
if not isinstance(layer_index, int):
|
|
40
|
-
raise SystemExit('Config "layer" must be an integer')
|
|
41
|
-
if proj not in ["q", "k", "v", "o", "all"]:
|
|
42
|
-
raise SystemExit('Config "proj" must be one of: q, k, v, o, all')
|
|
43
|
-
|
|
44
|
-
print(f"Loading model: {model_id}")
|
|
45
|
-
model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype=torch.float32, device_map="cpu")
|
|
46
|
-
|
|
47
|
-
layer = model.model.layers[layer_index]
|
|
48
|
-
attn = layer.self_attn
|
|
49
|
-
|
|
50
|
-
projections = {
|
|
51
|
-
"q": attn.q_proj,
|
|
52
|
-
"k": attn.k_proj,
|
|
53
|
-
"v": attn.v_proj,
|
|
54
|
-
"o": attn.o_proj,
|
|
55
|
-
}
|
|
56
|
-
|
|
57
|
-
projs_to_show = list(projections.keys()) if proj == "all" else [proj]
|
|
58
|
-
|
|
59
|
-
for name in projs_to_show:
|
|
60
|
-
proj = projections[name]
|
|
61
|
-
weight = proj.weight.data # [out_features, in_features]
|
|
62
|
-
|
|
63
|
-
print(f"\n{name.upper()}_proj weight shape: {weight.shape}")
|
|
64
|
-
print(f"{name.upper()}_proj weight[0, :8] (first row, first 8 cols):")
|
|
65
|
-
print(f" {weight[0, :8].tolist()}")
|
|
66
|
-
print(f"{name.upper()}_proj weight[:8, 0] (first 8 rows, first col):")
|
|
67
|
-
print(f" {weight[:8, 0].tolist()}")
|
|
68
|
-
|
|
69
|
-
# Some specific indices for comparison
|
|
70
|
-
if weight.shape[1] > 100:
|
|
71
|
-
print(f"{name.upper()}_proj weight[0, 100]: {weight[0, 100].item():.6f}")
|
|
72
|
-
|
|
73
|
-
|
|
74
|
-
if __name__ == "__main__":
|
|
75
|
-
main()
|
|
@@ -1,20 +0,0 @@
|
|
|
1
|
-
export declare const ERROR_CODES: {
|
|
2
|
-
readonly GPU_UNAVAILABLE: 'DOPPLER_GPU_UNAVAILABLE';
|
|
3
|
-
readonly GPU_DEVICE_FAILED: 'DOPPLER_GPU_DEVICE_FAILED';
|
|
4
|
-
readonly LOADER_MANIFEST_INVALID: 'DOPPLER_LOADER_MANIFEST_INVALID';
|
|
5
|
-
readonly LOADER_SHARD_INDEX_INVALID: 'DOPPLER_LOADER_SHARD_INDEX_INVALID';
|
|
6
|
-
readonly DISTRIBUTION_P2P_CONTRACT_UNSUPPORTED: 'DOPPLER_DISTRIBUTION_P2P_CONTRACT_UNSUPPORTED';
|
|
7
|
-
readonly DISTRIBUTION_P2P_TRANSPORT_UNCONFIGURED: 'DOPPLER_DISTRIBUTION_P2P_TRANSPORT_UNCONFIGURED';
|
|
8
|
-
readonly DISTRIBUTION_P2P_TRANSPORT_UNAVAILABLE: 'DOPPLER_DISTRIBUTION_P2P_TRANSPORT_UNAVAILABLE';
|
|
9
|
-
readonly DISTRIBUTION_P2P_TRANSPORT_TIMEOUT: 'DOPPLER_DISTRIBUTION_P2P_TRANSPORT_TIMEOUT';
|
|
10
|
-
readonly DISTRIBUTION_P2P_TRANSPORT_ABORTED: 'DOPPLER_DISTRIBUTION_P2P_TRANSPORT_ABORTED';
|
|
11
|
-
readonly DISTRIBUTION_P2P_TRANSPORT_INTEGRITY_MISMATCH: 'DOPPLER_DISTRIBUTION_P2P_TRANSPORT_INTEGRITY_MISMATCH';
|
|
12
|
-
readonly DISTRIBUTION_P2P_TRANSPORT_POLICY_DENIED: 'DOPPLER_DISTRIBUTION_P2P_TRANSPORT_POLICY_DENIED';
|
|
13
|
-
readonly DISTRIBUTION_P2P_TRANSPORT_INTERNAL: 'DOPPLER_DISTRIBUTION_P2P_TRANSPORT_INTERNAL';
|
|
14
|
-
readonly DISTRIBUTION_P2P_TRANSPORT_PAYLOAD_INVALID: 'DOPPLER_DISTRIBUTION_P2P_TRANSPORT_PAYLOAD_INVALID';
|
|
15
|
-
readonly DISTRIBUTION_SHARD_HASH_MISMATCH: 'DOPPLER_DISTRIBUTION_SHARD_HASH_MISMATCH';
|
|
16
|
-
readonly DISTRIBUTION_SHARD_SIZE_MISMATCH: 'DOPPLER_DISTRIBUTION_SHARD_SIZE_MISMATCH';
|
|
17
|
-
readonly DISTRIBUTION_SHARD_MANIFEST_VERSION_SET_MISMATCH: 'DOPPLER_DISTRIBUTION_SHARD_MANIFEST_VERSION_SET_MISMATCH';
|
|
18
|
-
};
|
|
19
|
-
|
|
20
|
-
export declare function createDopplerError(code: string, message: string): Error & { code?: string };
|
|
@@ -1,14 +0,0 @@
|
|
|
1
|
-
/**
|
|
2
|
-
* DOPPLER Bridge Background Script
|
|
3
|
-
* Phase 3: Native Messaging Bridge
|
|
4
|
-
*
|
|
5
|
-
* Handles:
|
|
6
|
-
* - Connection to native host
|
|
7
|
-
* - Message routing between web page and native host
|
|
8
|
-
* - Binary protocol handling
|
|
9
|
-
* - Backpressure management
|
|
10
|
-
*
|
|
11
|
-
* @module bridge/extension/background
|
|
12
|
-
*/
|
|
13
|
-
|
|
14
|
-
export {};
|
|
@@ -1,168 +0,0 @@
|
|
|
1
|
-
|
|
2
|
-
|
|
3
|
-
import { log } from '../../debug/index.js';
|
|
4
|
-
|
|
5
|
-
// ============================================================================
|
|
6
|
-
// Constants
|
|
7
|
-
// ============================================================================
|
|
8
|
-
|
|
9
|
-
const NATIVE_HOST_NAME = 'dev.reploid.doppler';
|
|
10
|
-
let nextPortCounter = 0;
|
|
11
|
-
|
|
12
|
-
function createPortId() {
|
|
13
|
-
const now = Date.now().toString(36);
|
|
14
|
-
nextPortCounter = (nextPortCounter + 1) >>> 0;
|
|
15
|
-
return `${now}-${nextPortCounter.toString(36)}`;
|
|
16
|
-
}
|
|
17
|
-
|
|
18
|
-
// ============================================================================
|
|
19
|
-
// Global State
|
|
20
|
-
// ============================================================================
|
|
21
|
-
|
|
22
|
-
|
|
23
|
-
const connections = new Map();
|
|
24
|
-
|
|
25
|
-
|
|
26
|
-
const messageQueues = new Map();
|
|
27
|
-
|
|
28
|
-
// ============================================================================
|
|
29
|
-
// Connection Handling
|
|
30
|
-
// ============================================================================
|
|
31
|
-
|
|
32
|
-
function handleWebConnection(webPort) {
|
|
33
|
-
if (webPort.name !== 'doppler-bridge') {
|
|
34
|
-
log.warn('DopplerBridge', `Unknown connection: ${webPort.name}`);
|
|
35
|
-
return;
|
|
36
|
-
}
|
|
37
|
-
|
|
38
|
-
log.info('DopplerBridge', 'Web page connected');
|
|
39
|
-
|
|
40
|
-
const portId = createPortId();
|
|
41
|
-
let nativePort;
|
|
42
|
-
|
|
43
|
-
// Connect to native host
|
|
44
|
-
try {
|
|
45
|
-
nativePort = chrome.runtime.connectNative(NATIVE_HOST_NAME);
|
|
46
|
-
log.info('DopplerBridge', 'Connected to native host');
|
|
47
|
-
} catch (err) {
|
|
48
|
-
log.error('DopplerBridge', `Failed to connect to native host: ${err.message}`);
|
|
49
|
-
webPort.postMessage({
|
|
50
|
-
type: 'error',
|
|
51
|
-
message: `Failed to connect to native host: ${err.message}`,
|
|
52
|
-
});
|
|
53
|
-
return;
|
|
54
|
-
}
|
|
55
|
-
|
|
56
|
-
// Store connection
|
|
57
|
-
connections.set(portId, { webPort, nativePort });
|
|
58
|
-
|
|
59
|
-
// Handle messages from web page
|
|
60
|
-
webPort.onMessage.addListener((message) => {
|
|
61
|
-
handleWebMessage(portId, message);
|
|
62
|
-
});
|
|
63
|
-
|
|
64
|
-
// Handle messages from native host
|
|
65
|
-
nativePort.onMessage.addListener((message) => {
|
|
66
|
-
handleNativeMessage(portId, message);
|
|
67
|
-
});
|
|
68
|
-
|
|
69
|
-
// Handle web page disconnect
|
|
70
|
-
webPort.onDisconnect.addListener(() => {
|
|
71
|
-
log.info('DopplerBridge', 'Web page disconnected');
|
|
72
|
-
cleanupConnection(portId);
|
|
73
|
-
});
|
|
74
|
-
|
|
75
|
-
// Handle native host disconnect
|
|
76
|
-
nativePort.onDisconnect.addListener(() => {
|
|
77
|
-
const error = chrome.runtime.lastError;
|
|
78
|
-
log.warn('DopplerBridge', `Native host disconnected: ${error?.message || 'unknown'}`);
|
|
79
|
-
|
|
80
|
-
// Notify web page
|
|
81
|
-
try {
|
|
82
|
-
webPort.postMessage({
|
|
83
|
-
type: 'error',
|
|
84
|
-
message: error?.message || 'Native host disconnected',
|
|
85
|
-
});
|
|
86
|
-
} catch {
|
|
87
|
-
// Port already closed
|
|
88
|
-
}
|
|
89
|
-
|
|
90
|
-
cleanupConnection(portId);
|
|
91
|
-
});
|
|
92
|
-
}
|
|
93
|
-
|
|
94
|
-
|
|
95
|
-
chrome.runtime.onConnectExternal.addListener(handleWebConnection);
|
|
96
|
-
|
|
97
|
-
|
|
98
|
-
function handleWebMessage(portId, message) {
|
|
99
|
-
const conn = connections.get(portId);
|
|
100
|
-
if (!conn) {
|
|
101
|
-
log.error('DopplerBridge', `No connection for port: ${portId}`);
|
|
102
|
-
return;
|
|
103
|
-
}
|
|
104
|
-
|
|
105
|
-
if (message.type === 'binary') {
|
|
106
|
-
// Forward binary message to native host
|
|
107
|
-
// Chrome native messaging uses JSON, so we send as array
|
|
108
|
-
conn.nativePort.postMessage({
|
|
109
|
-
type: 'binary',
|
|
110
|
-
data: message.data,
|
|
111
|
-
});
|
|
112
|
-
} else if (message.type === 'ack') {
|
|
113
|
-
// Forward ACK to native host
|
|
114
|
-
conn.nativePort.postMessage({
|
|
115
|
-
type: 'ack',
|
|
116
|
-
reqId: message.reqId,
|
|
117
|
-
});
|
|
118
|
-
} else {
|
|
119
|
-
log.warn('DopplerBridge', `Unknown message type from web: ${message.type}`);
|
|
120
|
-
}
|
|
121
|
-
}
|
|
122
|
-
|
|
123
|
-
|
|
124
|
-
function handleNativeMessage(portId, message) {
|
|
125
|
-
const conn = connections.get(portId);
|
|
126
|
-
if (!conn) {
|
|
127
|
-
log.error('DopplerBridge', `No connection for port: ${portId}`);
|
|
128
|
-
return;
|
|
129
|
-
}
|
|
130
|
-
|
|
131
|
-
if (message.type === 'binary') {
|
|
132
|
-
// Forward binary message to web page
|
|
133
|
-
conn.webPort.postMessage({
|
|
134
|
-
type: 'binary',
|
|
135
|
-
data: message.data,
|
|
136
|
-
});
|
|
137
|
-
} else if (message.type === 'error') {
|
|
138
|
-
conn.webPort.postMessage({
|
|
139
|
-
type: 'error',
|
|
140
|
-
message: message.message,
|
|
141
|
-
});
|
|
142
|
-
} else {
|
|
143
|
-
log.warn('DopplerBridge', `Unknown message type from native: ${message.type}`);
|
|
144
|
-
}
|
|
145
|
-
}
|
|
146
|
-
|
|
147
|
-
|
|
148
|
-
function cleanupConnection(portId) {
|
|
149
|
-
const conn = connections.get(portId);
|
|
150
|
-
if (conn) {
|
|
151
|
-
try {
|
|
152
|
-
conn.nativePort?.disconnect();
|
|
153
|
-
} catch {
|
|
154
|
-
// Already disconnected
|
|
155
|
-
}
|
|
156
|
-
connections.delete(portId);
|
|
157
|
-
}
|
|
158
|
-
messageQueues.delete(portId);
|
|
159
|
-
}
|
|
160
|
-
|
|
161
|
-
|
|
162
|
-
chrome.runtime.onConnect.addListener((port) => {
|
|
163
|
-
if (port.name === 'doppler-bridge') {
|
|
164
|
-
handleWebConnection(port);
|
|
165
|
-
}
|
|
166
|
-
});
|
|
167
|
-
|
|
168
|
-
log.info('DopplerBridge', 'Background script loaded');
|
|
@@ -1,34 +0,0 @@
|
|
|
1
|
-
{
|
|
2
|
-
"manifest_version": 3,
|
|
3
|
-
"name": "DOPPLER Bridge",
|
|
4
|
-
"version": "1.0.0",
|
|
5
|
-
"description": "Native bridge for DOPPLER LLM inference - bypasses browser storage limits",
|
|
6
|
-
|
|
7
|
-
"permissions": [
|
|
8
|
-
"nativeMessaging",
|
|
9
|
-
"storage"
|
|
10
|
-
],
|
|
11
|
-
|
|
12
|
-
"host_permissions": [],
|
|
13
|
-
|
|
14
|
-
"background": {
|
|
15
|
-
"service_worker": "background.js",
|
|
16
|
-
"type": "module"
|
|
17
|
-
},
|
|
18
|
-
|
|
19
|
-
"externally_connectable": {
|
|
20
|
-
"matches": [
|
|
21
|
-
"http://localhost:*/*",
|
|
22
|
-
"https://localhost:*/*",
|
|
23
|
-
"https://*.reploid.dev/*"
|
|
24
|
-
]
|
|
25
|
-
},
|
|
26
|
-
|
|
27
|
-
"icons": {
|
|
28
|
-
"16": "icons/icon16.png",
|
|
29
|
-
"48": "icons/icon48.png",
|
|
30
|
-
"128": "icons/icon128.png"
|
|
31
|
-
},
|
|
32
|
-
|
|
33
|
-
"key": "MIIBIjANBgkqhkiG9w0BAQEFAAOCAQ8AMIIBCgKCAQEA"
|
|
34
|
-
}
|