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
|
@@ -1,31 +0,0 @@
|
|
|
1
|
-
import type { ManifestEmbeddingPostprocessorSchema, TensorRole } from '../../config/schema/index.js';
|
|
2
|
-
|
|
3
|
-
export interface ParsedTransformerTensor {
|
|
4
|
-
name: string;
|
|
5
|
-
shape?: number[];
|
|
6
|
-
dtype?: string;
|
|
7
|
-
size?: number;
|
|
8
|
-
offset?: number;
|
|
9
|
-
sourcePath?: string;
|
|
10
|
-
role?: TensorRole;
|
|
11
|
-
group?: string | null;
|
|
12
|
-
}
|
|
13
|
-
|
|
14
|
-
export interface TransformerParserAdapter {
|
|
15
|
-
readJson: (suffix: string, label?: string) => Promise<Record<string, unknown>>;
|
|
16
|
-
fileExists: (suffix: string) => Promise<boolean>;
|
|
17
|
-
loadSingleSafetensors: (suffix: string) => Promise<ParsedTransformerTensor[]>;
|
|
18
|
-
loadShardedSafetensors: (indexJson: Record<string, unknown>) => Promise<ParsedTransformerTensor[]>;
|
|
19
|
-
}
|
|
20
|
-
|
|
21
|
-
export interface ParsedTransformerModel {
|
|
22
|
-
config: Record<string, unknown>;
|
|
23
|
-
generationConfig: Record<string, unknown> | null;
|
|
24
|
-
tensors: ParsedTransformerTensor[];
|
|
25
|
-
architectureHint: string;
|
|
26
|
-
embeddingPostprocessor: ManifestEmbeddingPostprocessorSchema | null;
|
|
27
|
-
}
|
|
28
|
-
|
|
29
|
-
export declare function parseTransformerModel(
|
|
30
|
-
adapter: TransformerParserAdapter
|
|
31
|
-
): Promise<ParsedTransformerModel>;
|
|
@@ -1,38 +0,0 @@
|
|
|
1
|
-
/**
|
|
2
|
-
* Quantization helpers shared across converter paths.
|
|
3
|
-
*
|
|
4
|
-
* @module converter/quantization-info
|
|
5
|
-
*/
|
|
6
|
-
|
|
7
|
-
import type { ConvertOptions } from './core.js';
|
|
8
|
-
import type { ConverterConfigSchema, QuantizationInfoSchema } from '../config/index.js';
|
|
9
|
-
|
|
10
|
-
export declare function normalizeQuantTag(value: string | null | undefined): string;
|
|
11
|
-
|
|
12
|
-
export declare function resolveManifestQuantization(quantize: string | null, fallback: string): string;
|
|
13
|
-
|
|
14
|
-
export declare function normalizeQ4KLayout(value: string | null | undefined): string | null;
|
|
15
|
-
|
|
16
|
-
export declare function buildQuantizationInfo(
|
|
17
|
-
opts: ConvertOptions | ConverterConfigSchema,
|
|
18
|
-
originalDtype: string,
|
|
19
|
-
embedDtype: string | null,
|
|
20
|
-
lmHeadDtype: string | null,
|
|
21
|
-
hasVision?: boolean,
|
|
22
|
-
hasAudio?: boolean,
|
|
23
|
-
hasProjector?: boolean,
|
|
24
|
-
modelConfig?: Record<string, unknown> | null
|
|
25
|
-
): QuantizationInfoSchema;
|
|
26
|
-
|
|
27
|
-
export declare function resolveModelId(
|
|
28
|
-
modelId: string | null,
|
|
29
|
-
baseName: string,
|
|
30
|
-
variantTag: string | undefined
|
|
31
|
-
): string;
|
|
32
|
-
|
|
33
|
-
export declare function toWebGPUDtype(dtype: string): string;
|
|
34
|
-
|
|
35
|
-
export declare function resolveEffectiveQuantizationInfo(
|
|
36
|
-
baseInfo: QuantizationInfoSchema,
|
|
37
|
-
tensors: Record<string, unknown>
|
|
38
|
-
): QuantizationInfoSchema;
|
|
@@ -1,16 +0,0 @@
|
|
|
1
|
-
import type { InferenceConfigSchema } from '../config/schema/inference.schema.js';
|
|
2
|
-
import type { ManifestInferenceSchema } from '../config/schema/manifest.schema.js';
|
|
3
|
-
|
|
4
|
-
/**
|
|
5
|
-
* Build RoPE configuration from explicit converter config and HF config.
|
|
6
|
-
*
|
|
7
|
-
* HF rope_scaling is treated as source of truth when present.
|
|
8
|
-
* If absent, Gemma 3 style rope_parameters are mapped from:
|
|
9
|
-
* - full_attention: ropeTheta + global scaling
|
|
10
|
-
* - sliding_attention: ropeLocalTheta + local scaling
|
|
11
|
-
* - proportional RoPE: explicit frequency-base dims for non-uniform head geometry
|
|
12
|
-
*/
|
|
13
|
-
export declare function buildRoPEConfig(
|
|
14
|
-
converterInference: InferenceConfigSchema,
|
|
15
|
-
config: Record<string, unknown>
|
|
16
|
-
): ManifestInferenceSchema['rope'];
|
|
@@ -1,12 +0,0 @@
|
|
|
1
|
-
export declare function resolveEosTokenId(options: {
|
|
2
|
-
config?: Record<string, unknown> | null;
|
|
3
|
-
generationConfig?: Record<string, unknown> | null;
|
|
4
|
-
tokenizer?: {
|
|
5
|
-
eosTokenId?: number;
|
|
6
|
-
eos_token_id?: number;
|
|
7
|
-
} | null;
|
|
8
|
-
tokenizerJson?: {
|
|
9
|
-
specialTokens?: { eos?: number; eos_token_id?: number };
|
|
10
|
-
special_tokens?: { eos?: number; eos_token_id?: number };
|
|
11
|
-
} | null;
|
|
12
|
-
}): number | number[];
|
|
@@ -1,114 +0,0 @@
|
|
|
1
|
-
# Reference Model Debug Scripts
|
|
2
|
-
|
|
3
|
-
Purpose: PyTorch-based reference checks for validating Doppler inference outputs.
|
|
4
|
-
|
|
5
|
-
## Scope
|
|
6
|
-
|
|
7
|
-
- Scripted comparisons for embeddings, attention, weights, and layer outputs.
|
|
8
|
-
- Debug workflow for isolating divergence sources.
|
|
9
|
-
|
|
10
|
-
These scripts run the original SafeTensor model weights via PyTorch/transformers to compare against DOPPLER's quantized inference. Use these to isolate whether bugs are in DOPPLER's kernels, weight loading, or architecture implementation. They are config-only and accept a single JSON config path (no flags).
|
|
11
|
-
|
|
12
|
-
## Setup
|
|
13
|
-
|
|
14
|
-
```bash
|
|
15
|
-
pip install torch transformers
|
|
16
|
-
```
|
|
17
|
-
|
|
18
|
-
## Scripts
|
|
19
|
-
|
|
20
|
-
### hf_embed_check.py
|
|
21
|
-
Compare embeddings and layer 0 outputs.
|
|
22
|
-
|
|
23
|
-
```json
|
|
24
|
-
// tmp-hf-embed.json
|
|
25
|
-
{
|
|
26
|
-
"model": "google/gemma-2-2b-it",
|
|
27
|
-
"prompt": "The color of the sky is"
|
|
28
|
-
}
|
|
29
|
-
```
|
|
30
|
-
|
|
31
|
-
```bash
|
|
32
|
-
python hf_embed_check.py ./tmp-hf-embed.json
|
|
33
|
-
```
|
|
34
|
-
|
|
35
|
-
### hf_attn_debug.py
|
|
36
|
-
Full attention debug: traces input_norm -> Q/K/V projections.
|
|
37
|
-
|
|
38
|
-
```json
|
|
39
|
-
// tmp-hf-attn.json
|
|
40
|
-
{
|
|
41
|
-
"model": "google/gemma-2-2b-it",
|
|
42
|
-
"prompt": "The color of the sky is",
|
|
43
|
-
"layer": 0
|
|
44
|
-
}
|
|
45
|
-
```
|
|
46
|
-
|
|
47
|
-
```bash
|
|
48
|
-
python hf_attn_debug.py ./tmp-hf-attn.json
|
|
49
|
-
```
|
|
50
|
-
|
|
51
|
-
### hf_weights.py
|
|
52
|
-
Dump Q/K/V/O projection weights for comparison.
|
|
53
|
-
|
|
54
|
-
```json
|
|
55
|
-
// tmp-hf-weights.json
|
|
56
|
-
{
|
|
57
|
-
"model": "google/gemma-2-2b-it",
|
|
58
|
-
"layer": 0,
|
|
59
|
-
"proj": "v"
|
|
60
|
-
}
|
|
61
|
-
```
|
|
62
|
-
|
|
63
|
-
```bash
|
|
64
|
-
python hf_weights.py ./tmp-hf-weights.json
|
|
65
|
-
```
|
|
66
|
-
|
|
67
|
-
### hf_rope_check.py
|
|
68
|
-
Verify RoPE frequencies and rotations.
|
|
69
|
-
|
|
70
|
-
```json
|
|
71
|
-
// tmp-hf-rope.json
|
|
72
|
-
{
|
|
73
|
-
"model": "google/gemma-2-2b-it",
|
|
74
|
-
"pos": 6,
|
|
75
|
-
"dim": 256,
|
|
76
|
-
"theta": 10000.0
|
|
77
|
-
}
|
|
78
|
-
```
|
|
79
|
-
|
|
80
|
-
```bash
|
|
81
|
-
python hf_rope_check.py ./tmp-hf-rope.json
|
|
82
|
-
```
|
|
83
|
-
|
|
84
|
-
### hf_layer_out.py
|
|
85
|
-
Compare hidden states at specific layers.
|
|
86
|
-
|
|
87
|
-
```json
|
|
88
|
-
// tmp-hf-layer.json
|
|
89
|
-
{
|
|
90
|
-
"model": "google/gemma-2-2b-it",
|
|
91
|
-
"prompt": "The color of the sky is",
|
|
92
|
-
"layers": [0, 12, 25],
|
|
93
|
-
"token": -1
|
|
94
|
-
}
|
|
95
|
-
```
|
|
96
|
-
|
|
97
|
-
```bash
|
|
98
|
-
python hf_layer_out.py ./tmp-hf-layer.json
|
|
99
|
-
```
|
|
100
|
-
|
|
101
|
-
## Debugging Process
|
|
102
|
-
|
|
103
|
-
1. **Identify divergence layer**: Run `hf_layer_out.py` to find where DOPPLER diverges from reference
|
|
104
|
-
2. **Check embeddings**: Run `hf_embed_check.py` to verify embedding scaling
|
|
105
|
-
3. **Debug attention**: Run `hf_attn_debug.py` at the diverging layer
|
|
106
|
-
4. **Check weights**: Run `hf_weights.py` to compare dequantized weight values
|
|
107
|
-
5. **Verify RoPE**: Run `hf_rope_check.py` if Q/K values diverge after position 0
|
|
108
|
-
|
|
109
|
-
## Common Issues
|
|
110
|
-
|
|
111
|
-
- **Embedding scale**: Gemma uses `sqrt(hidden_size)` scaling
|
|
112
|
-
- **RMSNorm offset**: Gemma uses `(1 + weight) * x` formula
|
|
113
|
-
- **Q4K dequant**: Check for repeated values (indexing bug)
|
|
114
|
-
- **Attention softcapping**: Gemma 2 uses `attn_logit_soft_capping=50`
|
|
@@ -1,114 +0,0 @@
|
|
|
1
|
-
#!/usr/bin/env python3
|
|
2
|
-
"""
|
|
3
|
-
Full attention debug: traces all intermediate values through layer 0.
|
|
4
|
-
|
|
5
|
-
Outputs: input_norm -> Q/K/V projections -> RoPE -> attention scores -> output
|
|
6
|
-
|
|
7
|
-
Usage:
|
|
8
|
-
python hf_attn_debug.py <config.json>
|
|
9
|
-
|
|
10
|
-
Config (JSON):
|
|
11
|
-
{
|
|
12
|
-
"model": "google/gemma-2-2b-it",
|
|
13
|
-
"prompt": "The color of the sky is",
|
|
14
|
-
"layer": 0
|
|
15
|
-
}
|
|
16
|
-
|
|
17
|
-
Requires: pip install torch transformers
|
|
18
|
-
"""
|
|
19
|
-
|
|
20
|
-
import json
|
|
21
|
-
import sys
|
|
22
|
-
import torch
|
|
23
|
-
from transformers import AutoModelForCausalLM, AutoTokenizer
|
|
24
|
-
|
|
25
|
-
|
|
26
|
-
def load_config():
|
|
27
|
-
if len(sys.argv) != 2:
|
|
28
|
-
raise SystemExit("Usage: python hf_attn_debug.py <config.json>")
|
|
29
|
-
path = sys.argv[1]
|
|
30
|
-
with open(path, "r", encoding="utf-8") as handle:
|
|
31
|
-
config = json.load(handle)
|
|
32
|
-
if not isinstance(config, dict):
|
|
33
|
-
raise SystemExit("Config must be a JSON object")
|
|
34
|
-
return config
|
|
35
|
-
|
|
36
|
-
|
|
37
|
-
def require_string(config, key):
|
|
38
|
-
value = config.get(key)
|
|
39
|
-
if not isinstance(value, str) or not value.strip():
|
|
40
|
-
raise SystemExit(f'Config "{key}" must be a non-empty string')
|
|
41
|
-
return value
|
|
42
|
-
|
|
43
|
-
|
|
44
|
-
def require_int(config, key):
|
|
45
|
-
value = config.get(key)
|
|
46
|
-
if not isinstance(value, int):
|
|
47
|
-
raise SystemExit(f'Config "{key}" must be an integer')
|
|
48
|
-
return value
|
|
49
|
-
|
|
50
|
-
|
|
51
|
-
def main():
|
|
52
|
-
config = load_config()
|
|
53
|
-
model_id = require_string(config, "model")
|
|
54
|
-
prompt = require_string(config, "prompt")
|
|
55
|
-
layer_index = require_int(config, "layer")
|
|
56
|
-
|
|
57
|
-
print(f"Loading model: {model_id}")
|
|
58
|
-
model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype=torch.float32, device_map="cpu")
|
|
59
|
-
tokenizer = AutoTokenizer.from_pretrained(model_id)
|
|
60
|
-
|
|
61
|
-
inputs = tokenizer(prompt, return_tensors="pt")
|
|
62
|
-
input_ids = inputs['input_ids']
|
|
63
|
-
num_tokens = input_ids.shape[1]
|
|
64
|
-
|
|
65
|
-
print(f"\nPrompt: {prompt}")
|
|
66
|
-
print(f"Token IDs: {input_ids[0].tolist()}")
|
|
67
|
-
print(f"Num tokens: {num_tokens}")
|
|
68
|
-
|
|
69
|
-
# Get model config
|
|
70
|
-
config = model.config
|
|
71
|
-
print(f"\nModel config:")
|
|
72
|
-
print(f" hidden_size: {config.hidden_size}")
|
|
73
|
-
print(f" num_attention_heads: {config.num_attention_heads}")
|
|
74
|
-
print(f" num_key_value_heads: {config.num_key_value_heads}")
|
|
75
|
-
print(f" head_dim: {config.head_dim}")
|
|
76
|
-
|
|
77
|
-
layer = model.model.layers[layer_index]
|
|
78
|
-
attn = layer.self_attn
|
|
79
|
-
|
|
80
|
-
# Get scaled embeddings as input
|
|
81
|
-
raw_embeddings = model.model.embed_tokens(input_ids)
|
|
82
|
-
hidden_states = raw_embeddings * (config.hidden_size ** 0.5)
|
|
83
|
-
print(f"\nInput hidden states (last token first 8): {hidden_states[0, -1, :8].tolist()}")
|
|
84
|
-
|
|
85
|
-
# Apply input norm
|
|
86
|
-
normed = layer.input_layernorm(hidden_states)
|
|
87
|
-
print(f"After input_norm (last token first 8): {normed[0, -1, :8].tolist()}")
|
|
88
|
-
print(f"After input_norm (token 0 first 8): {normed[0, 0, :8].tolist()}")
|
|
89
|
-
|
|
90
|
-
# Q, K, V projections
|
|
91
|
-
with torch.no_grad():
|
|
92
|
-
q = attn.q_proj(normed)
|
|
93
|
-
k = attn.k_proj(normed)
|
|
94
|
-
v = attn.v_proj(normed)
|
|
95
|
-
|
|
96
|
-
print(f"\nQ projection (last token first 8): {q[0, -1, :8].tolist()}")
|
|
97
|
-
print(f"K projection (last token first 8): {k[0, -1, :8].tolist()}")
|
|
98
|
-
print(f"V projection (last token first 8): {v[0, -1, :8].tolist()}")
|
|
99
|
-
|
|
100
|
-
print(f"\nQ projection (token 0 first 8): {q[0, 0, :8].tolist()}")
|
|
101
|
-
print(f"K projection (token 0 first 8): {k[0, 0, :8].tolist()}")
|
|
102
|
-
print(f"V projection (token 0 first 8): {v[0, 0, :8].tolist()}")
|
|
103
|
-
|
|
104
|
-
# Check for errors vs expected
|
|
105
|
-
print(f"\nPer-token maxAbs:")
|
|
106
|
-
for t in range(num_tokens):
|
|
107
|
-
q_max = q[0, t].abs().max().item()
|
|
108
|
-
k_max = k[0, t].abs().max().item()
|
|
109
|
-
v_max = v[0, t].abs().max().item()
|
|
110
|
-
print(f" Token {t}: Q_maxAbs={q_max:.4f}, K_maxAbs={k_max:.4f}, V_maxAbs={v_max:.4f}")
|
|
111
|
-
|
|
112
|
-
|
|
113
|
-
if __name__ == "__main__":
|
|
114
|
-
main()
|
|
@@ -1,89 +0,0 @@
|
|
|
1
|
-
#!/usr/bin/env python3
|
|
2
|
-
"""
|
|
3
|
-
Compare embeddings and layer outputs between HuggingFace and DOPPLER.
|
|
4
|
-
|
|
5
|
-
Usage:
|
|
6
|
-
python hf_embed_check.py <config.json>
|
|
7
|
-
|
|
8
|
-
Config (JSON):
|
|
9
|
-
{
|
|
10
|
-
"model": "google/gemma-2-2b-it",
|
|
11
|
-
"prompt": "The color of the sky is"
|
|
12
|
-
}
|
|
13
|
-
|
|
14
|
-
Requires: pip install torch transformers
|
|
15
|
-
"""
|
|
16
|
-
|
|
17
|
-
import json
|
|
18
|
-
import sys
|
|
19
|
-
import torch
|
|
20
|
-
from transformers import AutoModelForCausalLM, AutoTokenizer
|
|
21
|
-
|
|
22
|
-
|
|
23
|
-
def load_config():
|
|
24
|
-
if len(sys.argv) != 2:
|
|
25
|
-
raise SystemExit("Usage: python hf_embed_check.py <config.json>")
|
|
26
|
-
path = sys.argv[1]
|
|
27
|
-
with open(path, "r", encoding="utf-8") as handle:
|
|
28
|
-
config = json.load(handle)
|
|
29
|
-
if not isinstance(config, dict):
|
|
30
|
-
raise SystemExit("Config must be a JSON object")
|
|
31
|
-
return config
|
|
32
|
-
|
|
33
|
-
|
|
34
|
-
def require_string(config, key):
|
|
35
|
-
value = config.get(key)
|
|
36
|
-
if not isinstance(value, str) or not value.strip():
|
|
37
|
-
raise SystemExit(f'Config "{key}" must be a non-empty string')
|
|
38
|
-
return value
|
|
39
|
-
|
|
40
|
-
|
|
41
|
-
def main():
|
|
42
|
-
config = load_config()
|
|
43
|
-
model_id = require_string(config, "model")
|
|
44
|
-
prompt = require_string(config, "prompt")
|
|
45
|
-
|
|
46
|
-
print(f"Loading model: {model_id}")
|
|
47
|
-
model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype=torch.float32, device_map="cpu")
|
|
48
|
-
tokenizer = AutoTokenizer.from_pretrained(model_id)
|
|
49
|
-
|
|
50
|
-
inputs = tokenizer(prompt, return_tensors="pt")
|
|
51
|
-
input_ids = inputs['input_ids']
|
|
52
|
-
|
|
53
|
-
print(f"\nPrompt: {prompt}")
|
|
54
|
-
print(f"Token IDs: {input_ids[0].tolist()}")
|
|
55
|
-
|
|
56
|
-
# Get embeddings (raw, not scaled)
|
|
57
|
-
raw_embeddings = model.model.embed_tokens(input_ids)
|
|
58
|
-
print(f"\nRaw embeddings shape: {raw_embeddings.shape}")
|
|
59
|
-
print(f"Raw embeddings (last token first 5): {raw_embeddings[0, -1, :5].tolist()}")
|
|
60
|
-
|
|
61
|
-
# Gemma scales embeddings by sqrt(hidden_size)
|
|
62
|
-
hidden_size = model.config.hidden_size
|
|
63
|
-
scaled_embeddings = raw_embeddings * (hidden_size ** 0.5)
|
|
64
|
-
print(f"Scaled embeddings (last token first 5): {scaled_embeddings[0, -1, :5].tolist()}")
|
|
65
|
-
|
|
66
|
-
# Get layer 0 output via hooks
|
|
67
|
-
layer0_outputs = {}
|
|
68
|
-
|
|
69
|
-
def hook_layer0_output(module, input, output):
|
|
70
|
-
layer0_outputs['output'] = output[0].detach()
|
|
71
|
-
|
|
72
|
-
model.model.layers[0].register_forward_hook(hook_layer0_output)
|
|
73
|
-
|
|
74
|
-
# Run forward pass
|
|
75
|
-
with torch.no_grad():
|
|
76
|
-
outputs = model(input_ids, output_hidden_states=True)
|
|
77
|
-
|
|
78
|
-
print(f"\nLayer 0 output (last token first 5): {layer0_outputs['output'][0, -1, :5].tolist()}")
|
|
79
|
-
|
|
80
|
-
# Per-token comparison
|
|
81
|
-
print(f"\nPer-token maxAbs comparison:")
|
|
82
|
-
for t in range(inputs['input_ids'].shape[1]):
|
|
83
|
-
emb_max = scaled_embeddings[0, t].abs().max().item()
|
|
84
|
-
l0_max = layer0_outputs['output'][0, t].abs().max().item()
|
|
85
|
-
print(f" Token {t}: emb_maxAbs={emb_max:.2f}, L0_maxAbs={l0_max:.2f}")
|
|
86
|
-
|
|
87
|
-
|
|
88
|
-
if __name__ == "__main__":
|
|
89
|
-
main()
|
|
@@ -1,100 +0,0 @@
|
|
|
1
|
-
#!/usr/bin/env python3
|
|
2
|
-
"""
|
|
3
|
-
Compare layer outputs at each layer between HuggingFace and DOPPLER.
|
|
4
|
-
|
|
5
|
-
Usage:
|
|
6
|
-
python hf_layer_out.py <config.json>
|
|
7
|
-
|
|
8
|
-
Config (JSON):
|
|
9
|
-
{
|
|
10
|
-
"model": "google/gemma-2-2b-it",
|
|
11
|
-
"prompt": "The color of the sky is",
|
|
12
|
-
"layers": [0, 12, 25],
|
|
13
|
-
"token": -1
|
|
14
|
-
}
|
|
15
|
-
|
|
16
|
-
Requires: pip install torch transformers
|
|
17
|
-
"""
|
|
18
|
-
|
|
19
|
-
import json
|
|
20
|
-
import sys
|
|
21
|
-
import torch
|
|
22
|
-
from transformers import AutoModelForCausalLM, AutoTokenizer
|
|
23
|
-
|
|
24
|
-
|
|
25
|
-
def main():
|
|
26
|
-
if len(sys.argv) != 2:
|
|
27
|
-
raise SystemExit("Usage: python hf_layer_out.py <config.json>")
|
|
28
|
-
path = sys.argv[1]
|
|
29
|
-
with open(path, "r", encoding="utf-8") as handle:
|
|
30
|
-
config = json.load(handle)
|
|
31
|
-
if not isinstance(config, dict):
|
|
32
|
-
raise SystemExit("Config must be a JSON object")
|
|
33
|
-
|
|
34
|
-
model_id = config.get("model")
|
|
35
|
-
prompt = config.get("prompt")
|
|
36
|
-
layers = config.get("layers")
|
|
37
|
-
token = config.get("token")
|
|
38
|
-
|
|
39
|
-
if not isinstance(model_id, str) or not model_id.strip():
|
|
40
|
-
raise SystemExit('Config "model" must be a non-empty string')
|
|
41
|
-
if not isinstance(prompt, str) or not prompt.strip():
|
|
42
|
-
raise SystemExit('Config "prompt" must be a non-empty string')
|
|
43
|
-
if not isinstance(layers, list) or not layers:
|
|
44
|
-
raise SystemExit('Config "layers" must be a non-empty array')
|
|
45
|
-
if not all(isinstance(x, int) for x in layers):
|
|
46
|
-
raise SystemExit('Config "layers" entries must be integers')
|
|
47
|
-
if not isinstance(token, int):
|
|
48
|
-
raise SystemExit('Config "token" must be an integer')
|
|
49
|
-
|
|
50
|
-
layer_indices = layers
|
|
51
|
-
|
|
52
|
-
print(f"Loading model: {model_id}")
|
|
53
|
-
model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype=torch.float32, device_map="cpu")
|
|
54
|
-
tokenizer = AutoTokenizer.from_pretrained(model_id)
|
|
55
|
-
|
|
56
|
-
inputs = tokenizer(prompt, return_tensors="pt")
|
|
57
|
-
input_ids = inputs['input_ids']
|
|
58
|
-
num_tokens = input_ids.shape[1]
|
|
59
|
-
token_idx = token if token >= 0 else num_tokens + token
|
|
60
|
-
|
|
61
|
-
print(f"\nPrompt: {prompt}")
|
|
62
|
-
print(f"Token IDs: {input_ids[0].tolist()}")
|
|
63
|
-
print(f"Checking token index: {token_idx}")
|
|
64
|
-
|
|
65
|
-
# Run forward pass with hidden states
|
|
66
|
-
with torch.no_grad():
|
|
67
|
-
outputs = model(input_ids, output_hidden_states=True)
|
|
68
|
-
|
|
69
|
-
hidden_states = outputs.hidden_states # Tuple of [batch, seq, hidden]
|
|
70
|
-
print(f"\nNum hidden states: {len(hidden_states)} (embed + {len(hidden_states)-1} layers)")
|
|
71
|
-
|
|
72
|
-
for layer_idx in layer_indices:
|
|
73
|
-
if layer_idx >= len(hidden_states) - 1:
|
|
74
|
-
print(f"\nLayer {layer_idx}: OUT OF RANGE")
|
|
75
|
-
continue
|
|
76
|
-
|
|
77
|
-
# hidden_states[0] = embeddings, hidden_states[1] = layer 0 output, etc.
|
|
78
|
-
hs = hidden_states[layer_idx + 1] # +1 because index 0 is embeddings
|
|
79
|
-
|
|
80
|
-
vals = hs[0, token_idx, :8].tolist()
|
|
81
|
-
max_abs = hs[0, token_idx].abs().max().item()
|
|
82
|
-
mean_abs = hs[0, token_idx].abs().mean().item()
|
|
83
|
-
|
|
84
|
-
print(f"\nLayer {layer_idx} output (token {token_idx}):")
|
|
85
|
-
print(f" First 8: {[f'{v:.4f}' for v in vals]}")
|
|
86
|
-
print(f" maxAbs: {max_abs:.4f}, meanAbs: {mean_abs:.4f}")
|
|
87
|
-
|
|
88
|
-
# Also show BOS (token 0) for comparison
|
|
89
|
-
print(f"\n--- BOS Token (position 0) for reference ---")
|
|
90
|
-
for layer_idx in layer_indices:
|
|
91
|
-
if layer_idx >= len(hidden_states) - 1:
|
|
92
|
-
continue
|
|
93
|
-
hs = hidden_states[layer_idx + 1]
|
|
94
|
-
vals = hs[0, 0, :8].tolist()
|
|
95
|
-
max_abs = hs[0, 0].abs().max().item()
|
|
96
|
-
print(f"Layer {layer_idx} (BOS): first8={[f'{v:.4f}' for v in vals]}, maxAbs={max_abs:.4f}")
|
|
97
|
-
|
|
98
|
-
|
|
99
|
-
if __name__ == "__main__":
|
|
100
|
-
main()
|