doppler-gpu 0.4.3 → 0.4.5
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 +28 -0
- package/README.md +67 -34
- package/assets/doppler.svg +57 -1601
- package/package.json +36 -5
- package/src/cli/cli-output.js +34 -1
- package/src/cli/doppler-cli.d.ts +18 -0
- package/src/cli/doppler-cli.js +68 -35
- package/src/cli/doppler-serve.d.ts +210 -0
- package/src/cli/doppler-serve.js +468 -42
- package/src/client/doppler-registry.json +3 -5
- package/src/client/provider.d.ts +1 -1
- package/src/client/provider.js +1 -1
- package/src/client/runtime/index.js +13 -0
- package/src/client/runtime/lora.d.ts +11 -1
- package/src/client/runtime/lora.js +28 -7
- package/src/client/runtime/model-manager.js +35 -11
- package/src/client/runtime/model-session.d.ts +2 -1
- package/src/client/runtime/model-session.js +2 -2
- package/src/client/runtime/model-source.js +20 -1
- package/src/client/runtime/types.d.ts +8 -0
- package/src/client/wrap-pipeline-handle.d.ts +62 -0
- package/src/client/wrap-pipeline-handle.js +188 -13
- package/src/config/conversion/README.md +5 -0
- package/src/config/conversion/diffusiongemma/diffusiongemma-26b-a4b-it-q4k-ehf16-af16.json +646 -0
- package/src/config/conversion/embeddinggemma/google-embeddinggemma-300m-q4k-ehf16-af32.json +3 -3
- package/src/config/conversion/gemma3/gemma-3-1b-it-f16-af32.json +4 -4
- package/src/config/conversion/gemma3/gemma-3-1b-it-q4k-ehf16-af32.json +4 -4
- package/src/config/conversion/gemma3/gemma-3-270m-it-f16-af32.json +393 -0
- package/src/config/conversion/gemma3/gemma-3-270m-it-q4k-ehf16-af16.json +425 -0
- package/src/config/conversion/gemma3/gemma-3-270m-it-q4k-ehf16-af32.json +17 -9
- package/src/config/conversion/gemma3/translategemma-4b-1b-enes-q4k-ehf16-af32.json +4 -4
- package/src/config/conversion/gemma3/translategemma-4b-it-q4k-ehf16-af32.json +4 -4
- package/src/config/conversion/gemma4/gemma-4-12b-it-text-q4k-ehf16-af16.json +737 -0
- package/src/config/conversion/gemma4/gemma-4-12b-it-text-q4k-ehf16-af32.json +730 -0
- package/src/config/conversion/gemma4/gemma-4-12b-it-text-q4k-ehf16-hq4k-af16.json +780 -0
- package/src/config/conversion/gemma4/gemma-4-12b-it-text-w4a16-ct-ehf16-af16.json +716 -0
- package/src/config/conversion/gemma4/gemma-4-31b-it-text-q4k-ehf16-af16.json +5 -5
- package/src/config/conversion/gemma4/gemma-4-31b-it-text-q4k-ehf16-af32.json +4 -4
- package/src/config/conversion/gemma4/gemma-4-e2b-it-q4k-ehf16-af16-int4ple.json +14 -12
- package/src/config/conversion/gemma4/gemma-4-e2b-it-q4k-ehf16-af32-int4ple.json +7 -7
- package/src/config/conversion/gemma4/gemma-4-e2b-it-q4k-ehf16-af32.json +7 -7
- package/src/config/conversion/gemma4/gemma-4-moe-q4k-ehf16-af32.json +3 -3
- package/src/config/conversion/gpt-oss-20b-f16-xmxfp4.json +3 -3
- package/src/config/conversion/janus/janus-pro-1b-text-q4k-ehaf16.json +3 -3
- package/src/config/conversion/lfm2/lfm2.5-1.2b-instruct-q4k-ehf16-af32.json +3 -3
- package/src/config/conversion/qwen3/qwen-3-5-0-8b-q4k-ehaf16.json +8 -4
- package/src/config/conversion/qwen3/qwen-3-5-2b-q4k-ehaf16.json +8 -4
- package/src/config/conversion/qwen3/qwen-3-6-27b-q4k-eaf16.json +9 -5
- package/src/config/conversion/qwen3/qwen-3-6-27b-q4k-ehaf16.json +8 -4
- package/src/config/goal-completion-matrix.json +552 -0
- package/src/config/kernel-path-loader.js +192 -38
- package/src/config/kernels/kernel-ref-digests.js +52 -18
- package/src/config/kernels/registry.json +3017 -744
- package/src/config/merge-contract-check.js +7 -3
- package/src/config/merge-helpers.d.ts +2 -0
- package/src/config/merge-helpers.js +1 -0
- package/src/config/merge.d.ts +1 -0
- package/src/config/merge.js +30 -0
- package/src/config/model-lanes/catalog-lane-resolver.d.ts +66 -0
- package/src/config/model-lanes/catalog-lane-resolver.js +170 -0
- package/src/config/param-validator.js +4 -0
- package/src/config/runtime/experiments/verify/lfm2-verify.json +3 -3
- package/src/config/runtime/profiles/diffusiongemma-26b-a4b-throughput.json +26 -0
- package/src/config/runtime/profiles/gemma3-270m-q4k-throughput-overlapped-probe.json +42 -0
- package/src/config/runtime/profiles/gemma3-270m-q4k-throughput-probe.json +39 -0
- package/src/config/runtime/profiles/gemma4-12b-rdrr-safe-b4.json +48 -0
- package/src/config/runtime/profiles/gemma4-12b-rdrr-safe-single.json +46 -0
- package/src/config/schema/buffer-pool.schema.js +3 -3
- package/src/config/schema/claim-evidence-contract.schema.json +93 -0
- package/src/config/schema/command-surface-contract.schema.json +81 -0
- package/src/config/schema/converter.schema.d.ts +11 -0
- package/src/config/schema/converter.schema.js +8 -0
- package/src/config/schema/doppler.schema.d.ts +1 -0
- package/src/config/schema/doppler.schema.js +38 -6
- package/src/config/schema/execution-v1.schema.d.ts +27 -0
- package/src/config/schema/execution-v1.schema.js +28 -19
- package/src/config/schema/goal-completion-matrix.schema.json +223 -0
- package/src/config/schema/inference.schema.d.ts +11 -0
- package/src/config/schema/manifest.schema.d.ts +107 -1
- package/src/config/schema/manifest.schema.js +10 -1
- package/src/config/schema/model-artifact-contract.schema.json +119 -0
- package/src/config/schema/moe.schema.d.ts +13 -0
- package/src/config/schema/moe.schema.js +1 -1
- package/src/config/schema/policy-schema-registry.json +26 -0
- package/src/config/schema/subsystem-support-contract.schema.json +48 -0
- package/src/config/schema/training-metrics.schema.d.ts +1 -1
- package/src/config/schema/training-metrics.schema.js +4 -3
- package/src/config/source-packages/litert/gemma-4-12b-it.json +390 -0
- package/src/config/source-packages/litert/gemma-4-e2b-it.json +7 -7
- package/src/config/source-packages/litert/gemma-4-e4b-it.json +370 -0
- package/src/config/source-packages/registry.json +25 -0
- package/src/config/supported-operations.json +5 -0
- package/src/config/transforms/capability-transform-resolver.d.ts +13 -2
- package/src/config/transforms/execution-graph-transforms.d.ts +42 -0
- package/src/config/transforms/execution-graph-transforms.js +227 -230
- package/src/converter/conversion-plan.d.ts +2 -0
- package/src/converter/conversion-plan.js +18 -4
- package/src/converter/core.d.ts +8 -0
- package/src/converter/core.js +574 -12
- package/src/converter/quantization-info.js +179 -2
- package/src/converter/quantizer.d.ts +5 -4
- package/src/converter/quantizer.js +16 -2
- package/src/experimental/adapters/lora-loader.js +109 -24
- package/src/experimental/diffusion/index.d.ts +24 -0
- package/src/experimental/diffusion/index.js +10 -0
- package/src/experimental/distribution/descriptor-transport.d.ts +117 -0
- package/src/experimental/distribution/descriptor-transport.js +342 -0
- package/src/experimental/training/README.md +11 -0
- package/src/experimental/training/autograd.js +33 -8
- package/src/experimental/training/datasets/gepa-frontier.d.ts +41 -0
- package/src/experimental/training/datasets/gepa-frontier.js +220 -0
- package/src/experimental/training/datasets/index.d.ts +13 -1
- package/src/experimental/training/datasets/index.js +8 -1
- package/src/experimental/training/datasets/teacher-traces.d.ts +51 -0
- package/src/experimental/training/datasets/teacher-traces.js +279 -0
- package/src/experimental/training/datasets/text-pairs.d.ts +45 -0
- package/src/experimental/training/datasets/text-pairs.js +82 -5
- package/src/experimental/training/distillation/student-fixture.d.ts +5 -0
- package/src/experimental/training/distillation/student-fixture.js +290 -32
- package/src/experimental/training/export.d.ts +9 -0
- package/src/experimental/training/export.js +113 -7
- package/src/experimental/training/index.d.ts +13 -1
- package/src/experimental/training/index.js +13 -1
- package/src/experimental/training/lora-pipeline.d.ts +143 -0
- package/src/experimental/training/lora-pipeline.js +1737 -49
- package/src/experimental/training/lora.js +23 -1
- package/src/experimental/training/loss.js +1 -1
- package/src/experimental/training/operator-agent-eval.d.ts +65 -0
- package/src/experimental/training/operator-agent-eval.js +326 -0
- package/src/experimental/training/operator-command.js +310 -51
- package/src/experimental/training/runner.d.ts +1 -0
- package/src/experimental/training/runner.js +4 -0
- package/src/experimental/training/workload-packs/distill-glm52-doppler-agent-harness-gemma270m-f16-sft-tiny.json +160 -0
- package/src/experimental/training/workload-packs/distill-glm52-doppler-code-agent-gemma270m-f16-sft-tiny.json +175 -0
- package/src/experimental/training/workload-packs/distill-glm52-doppler-code-agent-sft-tiny.json +175 -0
- package/src/experimental/training/workload-packs/distill-glm52-doppler-js-json-gemma270m-f16-sft-tiny.json +160 -0
- package/src/experimental/training/workload-packs/distill-glm52-doppler-review-gemma270m-f16-sft-tiny.json +162 -0
- package/src/experimental/training/workload-packs/distill-glm52-doppler-wgsl-gemma270m-f16-sft-tiny.json +160 -0
- package/src/experimental/training/workload-packs/lora-doppler-agent-harness-gemma270m-f16-tiny.json +137 -0
- package/src/experimental/training/workload-packs/lora-doppler-code-agent-gemma270m-f16-tiny.json +152 -0
- package/src/experimental/training/workload-packs/lora-doppler-code-agent-tiny.json +152 -0
- package/src/experimental/training/workload-packs/lora-doppler-js-json-gemma270m-f16-tiny.json +137 -0
- package/src/experimental/training/workload-packs/lora-doppler-review-gemma270m-f16-tiny.json +139 -0
- package/src/experimental/training/workload-packs/lora-doppler-wgsl-gemma270m-f16-tiny.json +137 -0
- package/src/experimental/training/workload-packs/registry.json +186 -6
- package/src/experimental/training/workloads.d.ts +37 -0
- package/src/experimental/training/workloads.js +110 -1
- package/src/formats/rdrr/functional-descriptor.d.ts +73 -0
- package/src/formats/rdrr/functional-descriptor.js +230 -0
- package/src/formats/rdrr/index.d.ts +1 -0
- package/src/formats/rdrr/index.js +1 -0
- package/src/formats/rdrr/parsing.js +22 -6
- package/src/formats/rdrr/source-transform-contract.js +10 -1
- package/src/formats/rdrr/storage-descriptor.js +9 -2
- package/src/formats/rdrr/types.d.ts +5 -3
- package/src/formats/rdrr/validation.js +15 -1
- package/src/gpu/command-recorder.d.ts +43 -3
- package/src/gpu/command-recorder.js +122 -25
- package/src/gpu/device.js +111 -11
- package/src/gpu/kernel-tuner/cache.js +1 -5
- package/src/gpu/kernels/attention.js +17 -9
- package/src/gpu/kernels/attention_decode_online_head256_f16kv.wgsl +236 -0
- package/src/gpu/kernels/attention_streaming_f16.wgsl +7 -5
- package/src/gpu/kernels/backward/cross_entropy_backward.js +72 -3
- package/src/gpu/kernels/backward/cross_entropy_backward.wgsl +2 -2
- package/src/gpu/kernels/cast.js +74 -67
- package/src/gpu/kernels/codegen/patches/attention_streaming_f16.from.attention_streaming.diff +46 -1
- package/src/gpu/kernels/codegen/patches/gelu_f16.from.gelu.diff +0 -20
- package/src/gpu/kernels/codegen/patches/residual_f16.from.residual.diff +15 -12
- package/src/gpu/kernels/codegen/patches/rmsnorm_f16.from.rmsnorm.diff +59 -44
- package/src/gpu/kernels/diffusion-gemma-sampling.d.ts +19 -0
- package/src/gpu/kernels/diffusion-gemma-sampling.js +104 -0
- package/src/gpu/kernels/diffusion_gemma_sampling.wgsl +130 -0
- package/src/gpu/kernels/dispatch.js +13 -3
- package/src/gpu/kernels/fused_ffn.js +3 -4
- package/src/gpu/kernels/fused_matmul_litert_int4_f16a.wgsl +78 -0
- package/src/gpu/kernels/fused_matmul_litert_int4_f16a_f32out.wgsl +78 -0
- package/src/gpu/kernels/fused_matmul_litert_int4_f32a_f32out.wgsl +76 -0
- package/src/gpu/kernels/fused_matmul_rmsnorm.js +101 -80
- package/src/gpu/kernels/fused_matmul_w4a16_batched_f16a.wgsl +116 -0
- package/src/gpu/kernels/fused_matmul_w4a16_batched_f16a_f32out.wgsl +116 -0
- package/src/gpu/kernels/fused_matmul_w4a16_f16a.wgsl +95 -0
- package/src/gpu/kernels/fused_matmul_w4a16_f16a_f32out.wgsl +95 -0
- package/src/gpu/kernels/gather.d.ts +84 -1
- package/src/gpu/kernels/gather.js +353 -5
- package/src/gpu/kernels/gather_litert_int4_vec4_f16_out.wgsl +75 -0
- package/src/gpu/kernels/gather_litert_int4_vec4_f32_out.wgsl +72 -0
- package/src/gpu/kernels/gather_split4_f16_vec4_f16_out.wgsl +103 -0
- package/src/gpu/kernels/gather_split8_f16_vec4_f16_out.wgsl +143 -0
- package/src/gpu/kernels/gather_split8_f16_vec4_f32_out.wgsl +143 -0
- package/src/gpu/kernels/gelu.wgsl +11 -7
- package/src/gpu/kernels/gelu_f16.wgsl +11 -7
- package/src/gpu/kernels/gemma4-route-expert.d.ts +35 -0
- package/src/gpu/kernels/gemma4-route-expert.js +126 -0
- package/src/gpu/kernels/gemma4_route_q4_matmul_f16a.wgsl +210 -0
- package/src/gpu/kernels/index.d.ts +71 -0
- package/src/gpu/kernels/index.js +52 -0
- package/src/gpu/kernels/kv-cache-write.d.ts +22 -0
- package/src/gpu/kernels/kv-cache-write.js +126 -0
- package/src/gpu/kernels/kv_cache_write_f32_to_f16.wgsl +32 -0
- package/src/gpu/kernels/lm-head-argmax.d.ts +19 -0
- package/src/gpu/kernels/lm-head-argmax.js +218 -0
- package/src/gpu/kernels/lm_head_argmax_f16.wgsl +189 -0
- package/src/gpu/kernels/matmul-dispatch.d.ts +11 -0
- package/src/gpu/kernels/matmul-dispatch.js +37 -1
- package/src/gpu/kernels/matmul-selection.d.ts +5 -5
- package/src/gpu/kernels/matmul-selection.js +104 -9
- package/src/gpu/kernels/matmul.d.ts +1 -1
- package/src/gpu/kernels/matmul.js +151 -43
- package/src/gpu/kernels/moe.d.ts +1 -0
- package/src/gpu/kernels/moe.js +46 -18
- package/src/gpu/kernels/pipeline-cache.d.ts +8 -0
- package/src/gpu/kernels/pipeline-cache.js +24 -0
- package/src/gpu/kernels/rep-penalty.d.ts +21 -0
- package/src/gpu/kernels/rep-penalty.js +17 -9
- package/src/gpu/kernels/residual.d.ts +1 -0
- package/src/gpu/kernels/residual.js +13 -1
- package/src/gpu/kernels/residual.wgsl +3 -3
- package/src/gpu/kernels/residual_f16.wgsl +5 -3
- package/src/gpu/kernels/residual_f16_vec4.wgsl +6 -6
- package/src/gpu/kernels/residual_vec4.wgsl +6 -6
- package/src/gpu/kernels/rmsnorm-pair.d.ts +39 -0
- package/src/gpu/kernels/rmsnorm-pair.js +152 -0
- package/src/gpu/kernels/rmsnorm-qk.d.ts +41 -0
- package/src/gpu/kernels/rmsnorm-qk.js +128 -0
- package/src/gpu/kernels/rmsnorm.d.ts +31 -0
- package/src/gpu/kernels/rmsnorm.js +32 -7
- package/src/gpu/kernels/rmsnorm.wgsl +10 -6
- package/src/gpu/kernels/rmsnorm_f16.wgsl +7 -3
- package/src/gpu/kernels/rmsnorm_pair.wgsl +224 -0
- package/src/gpu/kernels/rmsnorm_qk.wgsl +118 -0
- package/src/gpu/kernels/rmsnorm_qk_f16.wgsl +143 -0
- package/src/gpu/kernels/rope-qk.d.ts +41 -0
- package/src/gpu/kernels/rope-qk.js +143 -0
- package/src/gpu/kernels/rope_qk.wgsl +95 -0
- package/src/gpu/kernels/rope_qk_f16.wgsl +97 -0
- package/src/gpu/kernels/rule-registry.js +50 -10
- package/src/gpu/kernels/sample.js +212 -187
- package/src/gpu/kernels/scatter_add_dynamic_expert_scale.wgsl +47 -0
- package/src/gpu/kernels/scatter_add_dynamic_f16_expert_scale.wgsl +49 -0
- package/src/gpu/kernels/scatter_add_dynamic_f16_weights_expert_scale.wgsl +49 -0
- package/src/gpu/kernels/scatter_add_routes_f16_weights_expert_scale.wgsl +45 -0
- package/src/gpu/kernels/shader-cache.js +1 -21
- package/src/gpu/kernels/soft-embedding.d.ts +28 -0
- package/src/gpu/kernels/soft-embedding.js +447 -0
- package/src/gpu/kernels/soft_embedding_logits_f16.wgsl +107 -0
- package/src/gpu/kernels/soft_embedding_split_f16.wgsl +46 -0
- package/src/gpu/kernels/softmax.wgsl +5 -80
- package/src/gpu/kernels/split-qkv-rmsnorm-qk.d.ts +44 -0
- package/src/gpu/kernels/split-qkv-rmsnorm-qk.js +154 -0
- package/src/gpu/kernels/split-qkv-rmsnorm-rope-qk.d.ts +59 -0
- package/src/gpu/kernels/split-qkv-rmsnorm-rope-qk.js +245 -0
- package/src/gpu/kernels/split_qkv_rmsnorm_qk.wgsl +146 -0
- package/src/gpu/kernels/split_qkv_rmsnorm_rope_qk.wgsl +185 -0
- package/src/gpu/kernels/split_qkv_rmsnorm_rope_qk_f16kv_cache.wgsl +187 -0
- package/src/gpu/kernels/uniform-utils.js +2 -5
- package/src/gpu/kernels/utils.d.ts +16 -1
- package/src/gpu/kernels/utils.js +42 -13
- package/src/gpu/readback-utils.js +10 -2
- package/src/gpu/uniform-cache.d.ts +10 -1
- package/src/gpu/uniform-cache.js +25 -7
- package/src/gpu/weight-buffer.d.ts +76 -7
- package/src/gpu/weight-buffer.js +85 -1
- package/src/inference/browser-harness-diffusion-energy-suites.js +269 -1
- package/src/inference/browser-harness-suite-helpers.d.ts +9 -0
- package/src/inference/browser-harness-suite-helpers.js +438 -2
- package/src/inference/browser-harness-text-helpers.d.ts +77 -0
- package/src/inference/browser-harness-text-helpers.js +236 -11
- package/src/inference/browser-harness.d.ts +8 -0
- package/src/inference/browser-harness.js +292 -8
- package/src/inference/decode-ring.d.ts +1 -0
- package/src/inference/decode-ring.js +7 -5
- package/src/inference/kv-cache/base.d.ts +25 -0
- package/src/inference/kv-cache/base.js +109 -1
- package/src/inference/kv-cache/mixed-geometry.d.ts +10 -0
- package/src/inference/kv-cache/mixed-geometry.js +109 -2
- package/src/inference/kv-cache/sliding-window.d.ts +14 -0
- package/src/inference/kv-cache/sliding-window.js +88 -0
- package/src/inference/model-lanes/execution-lane-resolver.d.ts +61 -0
- package/src/inference/model-lanes/execution-lane-resolver.js +126 -0
- package/src/inference/moe-router.d.ts +14 -1
- package/src/inference/moe-router.js +24 -1
- package/src/inference/pipelines/audio/gemma4.js +52 -1
- package/src/inference/pipelines/diffusion-gemma/config.d.ts +27 -0
- package/src/inference/pipelines/diffusion-gemma/config.js +149 -0
- package/src/inference/pipelines/diffusion-gemma/index.d.ts +31 -0
- package/src/inference/pipelines/diffusion-gemma/index.js +12 -0
- package/src/inference/pipelines/diffusion-gemma/pipeline.d.ts +69 -0
- package/src/inference/pipelines/diffusion-gemma/pipeline.js +380 -0
- package/src/inference/pipelines/diffusion-gemma/sampling.d.ts +128 -0
- package/src/inference/pipelines/diffusion-gemma/sampling.js +395 -0
- package/src/inference/pipelines/text/attention/dispatch-params.d.ts +28 -4
- package/src/inference/pipelines/text/attention/dispatch-params.js +176 -2
- package/src/inference/pipelines/text/attention/projections.d.ts +71 -2
- package/src/inference/pipelines/text/attention/projections.js +329 -49
- package/src/inference/pipelines/text/attention/quant-options.d.ts +38 -0
- package/src/inference/pipelines/text/attention/record.js +446 -73
- package/src/inference/pipelines/text/attention/run.js +251 -49
- package/src/inference/pipelines/text/attention/types.d.ts +5 -1
- package/src/inference/pipelines/text/config.d.ts +22 -1
- package/src/inference/pipelines/text/config.js +199 -22
- package/src/inference/pipelines/text/dtype-contract.d.ts +1 -1
- package/src/inference/pipelines/text/dtype-contract.js +4 -1
- package/src/inference/pipelines/text/embed.d.ts +4 -3
- package/src/inference/pipelines/text/embed.js +16 -6
- package/src/inference/pipelines/text/execution-plan.d.ts +8 -0
- package/src/inference/pipelines/text/execution-plan.js +6 -0
- package/src/inference/pipelines/text/execution-runtime-builders.d.ts +4 -0
- package/src/inference/pipelines/text/execution-runtime-builders.js +44 -28
- package/src/inference/pipelines/text/execution-v1.d.ts +2 -0
- package/src/inference/pipelines/text/execution-v1.js +373 -139
- package/src/inference/pipelines/text/ffn/dense.js +44 -13
- package/src/inference/pipelines/text/ffn/moe.d.ts +4 -1
- package/src/inference/pipelines/text/ffn/moe.js +9 -2
- package/src/inference/pipelines/text/ffn/sandwich.d.ts +3 -1
- package/src/inference/pipelines/text/ffn/sandwich.js +237 -15
- package/src/inference/pipelines/text/ffn/standard.d.ts +3 -1
- package/src/inference/pipelines/text/ffn/standard.js +44 -4
- package/src/inference/pipelines/text/generator-decode-policy.d.ts +81 -0
- package/src/inference/pipelines/text/generator-decode-policy.js +2 -12
- package/src/inference/pipelines/text/generator-helpers.js +5 -2
- package/src/inference/pipelines/text/generator-prefill-helpers.d.ts +84 -0
- package/src/inference/pipelines/text/generator-prefill-helpers.js +67 -1
- package/src/inference/pipelines/text/generator-runtime.d.ts +1 -0
- package/src/inference/pipelines/text/generator-steps.d.ts +38 -3
- package/src/inference/pipelines/text/generator-steps.js +268 -77
- package/src/inference/pipelines/text/generator.d.ts +39 -0
- package/src/inference/pipelines/text/generator.js +1105 -18
- package/src/inference/pipelines/text/init.d.ts +8 -1
- package/src/inference/pipelines/text/init.js +30 -1
- package/src/inference/pipelines/text/layer-plan-gpu.js +62 -14
- package/src/inference/pipelines/text/layer.js +272 -25
- package/src/inference/pipelines/text/linear-attention.js +6 -4
- package/src/inference/pipelines/text/logits/gpu.d.ts +49 -2
- package/src/inference/pipelines/text/logits/gpu.js +417 -66
- package/src/inference/pipelines/text/logits/index.d.ts +2 -2
- package/src/inference/pipelines/text/logits/index.js +96 -42
- package/src/inference/pipelines/text/logits/types.d.ts +2 -2
- package/src/inference/pipelines/text/lora-apply.js +17 -0
- package/src/inference/pipelines/text/model-load.js +5 -1
- package/src/inference/pipelines/text/moe-gpu.d.ts +13 -0
- package/src/inference/pipelines/text/moe-gpu.js +445 -16
- package/src/inference/pipelines/text/moe-impl.d.ts +13 -3
- package/src/inference/pipelines/text/moe-shape-validator.js +13 -1
- package/src/inference/pipelines/text/ops.d.ts +29 -2
- package/src/inference/pipelines/text/ops.js +19 -0
- package/src/inference/pipelines/text/per-layer-inputs.js +5 -0
- package/src/inference/pipelines/text/probes.js +4 -5
- package/src/inference/pipelines/text/sampling-config.d.ts +4 -10
- package/src/inference/pipelines/text/sampling-config.js +103 -64
- package/src/inference/pipelines/text/sampling.js +13 -9
- package/src/inference/pipelines/text/state.js +17 -0
- package/src/inference/pipelines/text/tsir-fixture-writer.d.ts +57 -0
- package/src/inference/pipelines/text/tsir-fixture-writer.js +8 -9
- package/src/inference/pipelines/text/types.d.ts +84 -3
- package/src/inference/pipelines/text/weights.d.ts +4 -2
- package/src/inference/pipelines/text/weights.js +4 -5
- package/src/inference/pipelines/text.d.ts +24 -0
- package/src/inference/pipelines/text.js +162 -38
- package/src/inference/pipelines/vision/gemma4.js +20 -9
- package/src/inference/tokenizer.d.ts +30 -0
- package/src/inference/tokenizer.js +340 -164
- package/src/inference/tokenizers/bundled.js +71 -49
- package/src/loader/doppler-loader.d.ts +11 -0
- package/src/loader/doppler-loader.js +290 -15
- package/src/loader/embedding-limit-preflight.d.ts +31 -0
- package/src/loader/embedding-limit-preflight.js +216 -0
- package/src/loader/embedding-loader.d.ts +8 -2
- package/src/loader/embedding-loader.js +263 -8
- package/src/loader/experts/expert-loader.d.ts +4 -1
- package/src/loader/experts/expert-loader.js +174 -5
- package/src/loader/final-weights-loader.d.ts +5 -0
- package/src/loader/final-weights-loader.js +167 -4
- package/src/loader/layer-loader.d.ts +2 -0
- package/src/loader/layer-loader.js +74 -6
- package/src/loader/loader-types.d.ts +59 -2
- package/src/loader/manifest-config.d.ts +8 -0
- package/src/loader/shard-resolver.js +37 -4
- package/src/loader/tensors/functional-descriptor-loader.d.ts +9 -0
- package/src/loader/tensors/functional-descriptor-loader.js +377 -0
- package/src/loader/tensors/tensor-loader.d.ts +36 -1
- package/src/loader/tensors/tensor-loader.js +437 -21
- package/src/loader/tensors/tensor-reader.d.ts +4 -1
- package/src/loader/tensors/tensor-reader.js +11 -4
- package/src/loader/weights.d.ts +3 -1
- package/src/memory/buffer-pool.js +12 -9
- package/src/models/diffusiongemma.d.ts +7 -0
- package/src/models/diffusiongemma.js +25 -0
- package/src/models/gemma4.js +24 -0
- package/src/models/qwen3.js +2 -2
- package/src/rules/execution-rules-contract-check.js +93 -10
- package/src/rules/inference/capability-transforms.rules.json +26 -2
- package/src/rules/inference/config.rules.json +1 -0
- package/src/rules/inference/dtype.rules.json +7 -0
- package/src/rules/inference/execution.rules.json +108 -17
- package/src/rules/inference/ffn.rules.json +18 -0
- package/src/rules/kernels/gather.rules.json +81 -0
- package/src/rules/kernels/matmul.rules.json +19 -0
- package/src/rules/kernels/moe.rules.json +3 -0
- package/src/rules/kernels/rep-penalty.rules.json +6 -0
- package/src/rules/kernels/rmsnorm-qk.rules.json +6 -0
- package/src/rules/kernels/rmsnorm.rules.json +42 -7
- package/src/rules/kernels/rope-qk.rules.json +6 -0
- package/src/rules/loader/tensor-loader.rules.json +8 -0
- package/src/rules/rule-registry.js +6 -0
- package/src/storage/artifact-storage-context.d.ts +1 -0
- package/src/storage/artifact-storage-context.js +10 -0
- package/src/tooling/browser-command-runner.d.ts +0 -3
- package/src/tooling/browser-command-runner.js +2 -14
- package/src/tooling/command-api-family-normalizers.js +4 -2
- package/src/tooling/command-api-helpers.js +2 -1
- package/src/tooling/command-api.d.ts +2 -0
- package/src/tooling/command-api.js +1 -0
- package/src/tooling/command-runner-shared.d.ts +4 -0
- package/src/tooling/command-runner-shared.js +2 -1
- package/src/tooling/diagnose-runner.js +2 -0
- package/src/tooling/litert-package-runtime.js +553 -16
- package/src/tooling/node-browser-command-runner.js +10 -4
- package/src/tooling/node-command-runner.js +7 -2
- package/src/tooling/node-convert-worker-pool.js +3 -0
- package/src/tooling/node-convert-worker.js +1 -0
- package/src/tooling/node-converter.js +12 -0
- package/src/tooling/node-webgpu.d.ts +1 -0
- package/src/tooling/node-webgpu.js +23 -11
- package/src/tooling/program-bundle.d.ts +52 -1
- package/src/tooling/rdrr-integrity-refresh.d.ts +18 -0
- package/src/tooling/source-artifact-adapter.js +7 -2
- package/src/tooling/source-package-profiles.d.ts +18 -1
- package/src/tooling/source-package-profiles.js +109 -1
- package/src/tooling/source-runtime-bundle.d.ts +1 -0
- package/src/tooling/source-runtime-bundle.js +27 -0
- package/src/tooling/source-runtime-converter-config.js +12 -3
- package/src/tooling-experimental-exports.shared.d.ts +10 -0
- package/src/tooling-experimental-exports.shared.js +10 -0
- package/src/version.js +1 -1
|
@@ -131,12 +131,11 @@
|
|
|
131
131
|
"qwen-3-5-0-8b-wq4k-ef16-hf16-f16"
|
|
132
132
|
],
|
|
133
133
|
"modes": [
|
|
134
|
-
"text"
|
|
135
|
-
"vision"
|
|
134
|
+
"text"
|
|
136
135
|
],
|
|
137
136
|
"hf": {
|
|
138
137
|
"repoId": "Clocksmith/rdrr",
|
|
139
|
-
"revision": "
|
|
138
|
+
"revision": "95a01447eecbf13fc5964986f507b08ded0cd40f",
|
|
140
139
|
"path": "models/qwen-3-5-0-8b-q4k-ehaf16"
|
|
141
140
|
}
|
|
142
141
|
},
|
|
@@ -155,8 +154,7 @@
|
|
|
155
154
|
"qwen-3-5-2b-wq4k-ef16-hf16-f16"
|
|
156
155
|
],
|
|
157
156
|
"modes": [
|
|
158
|
-
"text"
|
|
159
|
-
"vision"
|
|
157
|
+
"text"
|
|
160
158
|
],
|
|
161
159
|
"hf": {
|
|
162
160
|
"repoId": "Clocksmith/rdrr",
|
package/src/client/provider.d.ts
CHANGED
|
@@ -148,4 +148,4 @@ export interface DopplerProvider {
|
|
|
148
148
|
|
|
149
149
|
export function createDopplerProvider(config: ProviderConfig): DopplerProvider;
|
|
150
150
|
|
|
151
|
-
export { wrapPipelineAsHandle } from './wrap-pipeline-handle.js';
|
|
151
|
+
export { wrapPipelineAsHandle, wrapPipelineAsDreamProvider } from './wrap-pipeline-handle.js';
|
package/src/client/provider.js
CHANGED
|
@@ -8,7 +8,7 @@ import { isNodeRuntime } from '../utils/runtime-env.js';
|
|
|
8
8
|
|
|
9
9
|
// Re-export handle adapter so consumers can reach both via
|
|
10
10
|
// `doppler-gpu/provider` without a second subpath.
|
|
11
|
-
export { wrapPipelineAsHandle } from './wrap-pipeline-handle.js';
|
|
11
|
+
export { wrapPipelineAsHandle, wrapPipelineAsDreamProvider } from './wrap-pipeline-handle.js';
|
|
12
12
|
|
|
13
13
|
// Canonical four routing modes (30-hybrid-routing-sdk.md).
|
|
14
14
|
const VALID_POLICY_MODES = new Set([
|
|
@@ -14,6 +14,8 @@ import {
|
|
|
14
14
|
createNodeFileArtifactStorageContext,
|
|
15
15
|
} from '../../storage/artifact-storage-context.js';
|
|
16
16
|
import { isNodeRuntime } from '../../utils/runtime-env.js';
|
|
17
|
+
import { resolveManifestGpuResidentEmbeddingLimitError } from '../../loader/embedding-limit-preflight.js';
|
|
18
|
+
import { initDevice } from '../../gpu/device.js';
|
|
17
19
|
|
|
18
20
|
function emitLoadProgress(callback, phase, percent, message) {
|
|
19
21
|
if (typeof callback !== 'function') return;
|
|
@@ -94,6 +96,17 @@ export function createDopplerRuntimeService({
|
|
|
94
96
|
? { text: JSON.stringify(resolved.manifest), manifest: resolved.manifest }
|
|
95
97
|
: await fetchManifestPayloadFromBaseUrl(resolved.baseUrl);
|
|
96
98
|
const resolvedArtifactSource = await resolveManifestArtifactSource(resolved, manifestPayload);
|
|
99
|
+
await initDevice();
|
|
100
|
+
const embeddingLimitError = resolveManifestGpuResidentEmbeddingLimitError(
|
|
101
|
+
resolvedArtifactSource.manifest,
|
|
102
|
+
{
|
|
103
|
+
storageManifest: resolvedArtifactSource.storageManifest,
|
|
104
|
+
runtimeConfig: options.runtimeConfig,
|
|
105
|
+
}
|
|
106
|
+
);
|
|
107
|
+
if (embeddingLimitError) {
|
|
108
|
+
throw embeddingLimitError;
|
|
109
|
+
}
|
|
97
110
|
const cachedResolved = manifestPayload.manifest?.weightsRef == null
|
|
98
111
|
? await resolveCachedNodeQuickstartSource(
|
|
99
112
|
resolvedArtifactSource,
|
|
@@ -2,9 +2,19 @@ import type { RDRRManifest } from '../../formats/rdrr/index.js';
|
|
|
2
2
|
import type { InferencePipeline } from '../../inference/pipelines/text.js';
|
|
3
3
|
import type { LoRAManifest } from './types.js';
|
|
4
4
|
|
|
5
|
+
export interface LoRALoadOptions {
|
|
6
|
+
basePath?: string;
|
|
7
|
+
resolvePath?: (path: string) => string;
|
|
8
|
+
readFile?: (path: string) => Promise<ArrayBuffer | Uint8Array>;
|
|
9
|
+
readOPFS?: (path: string) => Promise<ArrayBuffer | Uint8Array>;
|
|
10
|
+
writeOPFS?: (path: string, data: ArrayBuffer | Uint8Array) => Promise<void>;
|
|
11
|
+
fetchUrl?: (url: string) => Promise<ArrayBuffer | Uint8Array>;
|
|
12
|
+
}
|
|
13
|
+
|
|
5
14
|
export declare function loadLoRAAdapterForPipeline(
|
|
6
15
|
pipeline: InferencePipeline | null | undefined,
|
|
7
|
-
adapter: LoRAManifest | RDRRManifest | string
|
|
16
|
+
adapter: LoRAManifest | RDRRManifest | string,
|
|
17
|
+
loadOptions?: LoRALoadOptions
|
|
8
18
|
): Promise<void>;
|
|
9
19
|
|
|
10
20
|
export declare function activateLoRAFromTrainingOutputForPipeline(
|
|
@@ -9,11 +9,12 @@ async function getExperimentalLoRAModule() {
|
|
|
9
9
|
return loraModulePromise;
|
|
10
10
|
}
|
|
11
11
|
|
|
12
|
-
function createLoRALoadOptions() {
|
|
12
|
+
function createLoRALoadOptions(overrides = {}) {
|
|
13
13
|
return {
|
|
14
14
|
readOPFS: readOPFSFile,
|
|
15
15
|
writeOPFS: writeOPFSFile,
|
|
16
16
|
fetchUrl: fetchArrayBuffer,
|
|
17
|
+
...overrides,
|
|
17
18
|
};
|
|
18
19
|
}
|
|
19
20
|
|
|
@@ -22,12 +23,12 @@ function getActiveLoRAName(pipeline) {
|
|
|
22
23
|
return active ? active.name : null;
|
|
23
24
|
}
|
|
24
25
|
|
|
25
|
-
export async function loadLoRAAdapterForPipeline(pipeline, adapter) {
|
|
26
|
+
export async function loadLoRAAdapterForPipeline(pipeline, adapter, loadOptions = {}) {
|
|
26
27
|
if (!pipeline) {
|
|
27
28
|
throw new Error('No model loaded. Call load() first.');
|
|
28
29
|
}
|
|
29
30
|
|
|
30
|
-
const options = createLoRALoadOptions();
|
|
31
|
+
const options = createLoRALoadOptions(loadOptions);
|
|
31
32
|
let lora;
|
|
32
33
|
if (typeof adapter === 'string') {
|
|
33
34
|
const { loadLoRAFromUrl } = await getExperimentalLoRAModule();
|
|
@@ -45,6 +46,28 @@ export async function loadLoRAAdapterForPipeline(pipeline, adapter) {
|
|
|
45
46
|
log.info('doppler', `LoRA adapter loaded: ${lora.name}`);
|
|
46
47
|
}
|
|
47
48
|
|
|
49
|
+
async function readLocalJson(path) {
|
|
50
|
+
const { readFile } = await import('node:fs/promises');
|
|
51
|
+
const raw = await readFile(path, 'utf8');
|
|
52
|
+
return JSON.parse(raw);
|
|
53
|
+
}
|
|
54
|
+
|
|
55
|
+
async function createLocalFileLoadOptions(path) {
|
|
56
|
+
const { readFile } = await import('node:fs/promises');
|
|
57
|
+
const { dirname, isAbsolute, join } = await import('node:path');
|
|
58
|
+
const basePath = dirname(path);
|
|
59
|
+
return {
|
|
60
|
+
basePath,
|
|
61
|
+
resolvePath(filePath) {
|
|
62
|
+
return isAbsolute(filePath) ? filePath : join(basePath, filePath);
|
|
63
|
+
},
|
|
64
|
+
async readFile(filePath) {
|
|
65
|
+
const data = await readFile(filePath);
|
|
66
|
+
return data.buffer.slice(data.byteOffset, data.byteOffset + data.byteLength);
|
|
67
|
+
},
|
|
68
|
+
};
|
|
69
|
+
}
|
|
70
|
+
|
|
48
71
|
export async function activateLoRAFromTrainingOutputForPipeline(pipeline, trainingOutput) {
|
|
49
72
|
if (!pipeline) {
|
|
50
73
|
return {
|
|
@@ -124,10 +147,8 @@ export async function activateLoRAFromTrainingOutputForPipeline(pipeline, traini
|
|
|
124
147
|
if (!isNode) {
|
|
125
148
|
throw new Error('adapterManifestPath local files require Node runtime.');
|
|
126
149
|
}
|
|
127
|
-
const
|
|
128
|
-
|
|
129
|
-
const manifest = JSON.parse(raw);
|
|
130
|
-
await loadLoRAAdapterForPipeline(pipeline, manifest);
|
|
150
|
+
const manifest = await readLocalJson(path);
|
|
151
|
+
await loadLoRAAdapterForPipeline(pipeline, manifest, await createLocalFileLoadOptions(path));
|
|
131
152
|
return {
|
|
132
153
|
activated: true,
|
|
133
154
|
adapterName: getActiveLoRAName(pipeline),
|
|
@@ -582,16 +582,21 @@ export async function loadModel(modelId, modelUrl = null, onProgress = null, loc
|
|
|
582
582
|
}
|
|
583
583
|
}
|
|
584
584
|
|
|
585
|
-
|
|
586
|
-
|
|
587
|
-
throw new Error('No model loaded. Call loadModel() first.');
|
|
588
|
-
}
|
|
589
|
-
|
|
590
|
-
const options = {
|
|
585
|
+
function createLoRALoadOptions(overrides = {}) {
|
|
586
|
+
return {
|
|
591
587
|
readOPFS: readOPFSFile,
|
|
592
588
|
writeOPFS: writeOPFSFile,
|
|
593
589
|
fetchUrl: fetchArrayBuffer,
|
|
590
|
+
...overrides,
|
|
594
591
|
};
|
|
592
|
+
}
|
|
593
|
+
|
|
594
|
+
async function loadLoRAAdapter(adapter, loadOptions = {}) {
|
|
595
|
+
if (!pipeline) {
|
|
596
|
+
throw new Error('No model loaded. Call loadModel() first.');
|
|
597
|
+
}
|
|
598
|
+
|
|
599
|
+
const options = createLoRALoadOptions(loadOptions);
|
|
595
600
|
|
|
596
601
|
let lora;
|
|
597
602
|
if (typeof adapter === 'string') {
|
|
@@ -610,6 +615,28 @@ async function loadLoRAAdapter(adapter) {
|
|
|
610
615
|
log.info('DopplerProvider', `LoRA adapter loaded: ${lora.name}`);
|
|
611
616
|
}
|
|
612
617
|
|
|
618
|
+
async function readLocalJson(path) {
|
|
619
|
+
const { readFile } = await import('node:fs/promises');
|
|
620
|
+
const raw = await readFile(path, 'utf8');
|
|
621
|
+
return JSON.parse(raw);
|
|
622
|
+
}
|
|
623
|
+
|
|
624
|
+
async function createLocalFileLoadOptions(path) {
|
|
625
|
+
const { readFile } = await import('node:fs/promises');
|
|
626
|
+
const { dirname, isAbsolute, join } = await import('node:path');
|
|
627
|
+
const basePath = dirname(path);
|
|
628
|
+
return {
|
|
629
|
+
basePath,
|
|
630
|
+
resolvePath(filePath) {
|
|
631
|
+
return isAbsolute(filePath) ? filePath : join(basePath, filePath);
|
|
632
|
+
},
|
|
633
|
+
async readFile(filePath) {
|
|
634
|
+
const data = await readFile(filePath);
|
|
635
|
+
return data.buffer.slice(data.byteOffset, data.byteOffset + data.byteLength);
|
|
636
|
+
},
|
|
637
|
+
};
|
|
638
|
+
}
|
|
639
|
+
|
|
613
640
|
export async function activateLoRAFromTrainingOutput(trainingOutput) {
|
|
614
641
|
if (!pipeline) {
|
|
615
642
|
return {
|
|
@@ -688,10 +715,8 @@ export async function activateLoRAFromTrainingOutput(trainingOutput) {
|
|
|
688
715
|
if (!isNode) {
|
|
689
716
|
throw new Error('adapterManifestPath local files require Node runtime.');
|
|
690
717
|
}
|
|
691
|
-
const
|
|
692
|
-
|
|
693
|
-
const manifest = JSON.parse(raw);
|
|
694
|
-
await loadLoRAAdapter(manifest);
|
|
718
|
+
const manifest = await readLocalJson(path);
|
|
719
|
+
await loadLoRAAdapter(manifest, await createLocalFileLoadOptions(path));
|
|
695
720
|
return {
|
|
696
721
|
activated: true,
|
|
697
722
|
adapterName: getActiveLoRA(),
|
|
@@ -722,4 +747,3 @@ export function getActiveLoRA() {
|
|
|
722
747
|
const active = pipeline?.getActiveLoRA() || null;
|
|
723
748
|
return active ? active.name : null;
|
|
724
749
|
}
|
|
725
|
-
|
|
@@ -4,6 +4,7 @@ import type { GenerateOptions } from '../../generation/index.js';
|
|
|
4
4
|
import type { RDRRManifest } from '../../formats/rdrr/index.js';
|
|
5
5
|
import type { LogitsStepResult, PrefillResult } from '../../inference/pipelines/text/types.d.ts';
|
|
6
6
|
import type { LoRAManifest } from './types.js';
|
|
7
|
+
import type { LoRALoadOptions } from './lora.js';
|
|
7
8
|
|
|
8
9
|
export type DopplerGenerateOptions = Omit<GenerateOptions, 'stopTokens'>;
|
|
9
10
|
|
|
@@ -21,7 +22,7 @@ export interface DopplerModelHandle {
|
|
|
21
22
|
generateText(prompt: string, options?: DopplerGenerateOptions): Promise<string>;
|
|
22
23
|
chat(messages: ChatMessage[], options?: DopplerGenerateOptions): AsyncGenerator<string, void, void>;
|
|
23
24
|
chatText(messages: ChatMessage[], options?: DopplerGenerateOptions): Promise<DopplerChatResponse>;
|
|
24
|
-
loadLoRA(adapter: LoRAManifest | RDRRManifest | string): Promise<void>;
|
|
25
|
+
loadLoRA(adapter: LoRAManifest | RDRRManifest | string, loadOptions?: LoRALoadOptions): Promise<void>;
|
|
25
26
|
activateLoRAFromTrainingOutput(
|
|
26
27
|
trainingOutput:
|
|
27
28
|
| string
|
|
@@ -106,8 +106,8 @@ export function createModelHandle(pipeline, resolved) {
|
|
|
106
106
|
return pipeline.manifest?.inference?.supportsVision === true
|
|
107
107
|
&& pipeline.visionCapable === true;
|
|
108
108
|
},
|
|
109
|
-
async loadLoRA(adapter) {
|
|
110
|
-
return loadLoRAAdapterForPipeline(pipeline, adapter);
|
|
109
|
+
async loadLoRA(adapter, loadOptions = {}) {
|
|
110
|
+
return loadLoRAAdapterForPipeline(pipeline, adapter, loadOptions);
|
|
111
111
|
},
|
|
112
112
|
async activateLoRAFromTrainingOutput(trainingOutput) {
|
|
113
113
|
return activateLoRAFromTrainingOutputForPipeline(pipeline, trainingOutput);
|
|
@@ -32,7 +32,26 @@ export function resolveLoadProgressHandlers(options = {}, defaultLoadProgressLog
|
|
|
32
32
|
}
|
|
33
33
|
|
|
34
34
|
export async function fetchManifestPayloadFromBaseUrl(baseUrl) {
|
|
35
|
-
const
|
|
35
|
+
const manifestUrl = getManifestUrl(baseUrl);
|
|
36
|
+
if (/^file:\/\//i.test(manifestUrl)) {
|
|
37
|
+
let text;
|
|
38
|
+
try {
|
|
39
|
+
const [{ readFile }, { fileURLToPath }] = await Promise.all([
|
|
40
|
+
import('node:fs/promises'),
|
|
41
|
+
import('node:url'),
|
|
42
|
+
]);
|
|
43
|
+
text = await readFile(fileURLToPath(manifestUrl), 'utf8');
|
|
44
|
+
} catch (error) {
|
|
45
|
+
throw new Error(
|
|
46
|
+
`Failed to read manifest from ${baseUrl}: ${error?.message || String(error)}`
|
|
47
|
+
);
|
|
48
|
+
}
|
|
49
|
+
return {
|
|
50
|
+
text,
|
|
51
|
+
manifest: parseManifest(text),
|
|
52
|
+
};
|
|
53
|
+
}
|
|
54
|
+
const response = await fetch(manifestUrl);
|
|
36
55
|
if (!response.ok) {
|
|
37
56
|
throw new Error(`Failed to fetch manifest from ${baseUrl}: ${response.status}`);
|
|
38
57
|
}
|
|
@@ -9,8 +9,16 @@ export interface ExtensionBridgeClient {
|
|
|
9
9
|
export interface LoRAManifest {
|
|
10
10
|
adapterType?: string;
|
|
11
11
|
modelType?: string;
|
|
12
|
+
id?: string;
|
|
12
13
|
name?: string;
|
|
14
|
+
baseModel?: string;
|
|
13
15
|
rank?: number;
|
|
16
|
+
alpha?: number;
|
|
17
|
+
targetModules?: string[];
|
|
18
|
+
weightsFormat?: string;
|
|
19
|
+
weightsPath?: string;
|
|
20
|
+
checksum?: string;
|
|
21
|
+
checksumAlgorithm?: string;
|
|
14
22
|
tensors?: unknown;
|
|
15
23
|
[key: string]: unknown;
|
|
16
24
|
}
|
|
@@ -18,3 +18,65 @@ export declare function wrapPipelineAsHandle(
|
|
|
18
18
|
deviceInfo?: Record<string, unknown>;
|
|
19
19
|
}
|
|
20
20
|
): WrappedPipelineHandle;
|
|
21
|
+
|
|
22
|
+
export interface DreamProviderAttachLoRARequest {
|
|
23
|
+
adapterId: string;
|
|
24
|
+
version?: string;
|
|
25
|
+
baseModel?: string;
|
|
26
|
+
baseModelId?: string;
|
|
27
|
+
rank: number;
|
|
28
|
+
alpha?: number;
|
|
29
|
+
scale?: number;
|
|
30
|
+
targetModules?: string[];
|
|
31
|
+
layers: Map<number, Record<string, {
|
|
32
|
+
a: Float32Array | number[] | ArrayLike<number>;
|
|
33
|
+
b: Float32Array | number[] | ArrayLike<number>;
|
|
34
|
+
}>>;
|
|
35
|
+
}
|
|
36
|
+
|
|
37
|
+
export interface DreamProviderGenerateRequest {
|
|
38
|
+
prompt: unknown;
|
|
39
|
+
loraAdapterId?: string | null;
|
|
40
|
+
samplingOptions?: Record<string, unknown>;
|
|
41
|
+
}
|
|
42
|
+
|
|
43
|
+
export interface DreamProviderGenerateResult {
|
|
44
|
+
text: string;
|
|
45
|
+
useLora: boolean;
|
|
46
|
+
baseModelId: string;
|
|
47
|
+
loraAdapterId: string | null;
|
|
48
|
+
}
|
|
49
|
+
|
|
50
|
+
export interface DreamCausalLmBaseProvider {
|
|
51
|
+
readonly modelId: string;
|
|
52
|
+
readonly manifest: Record<string, unknown> | null;
|
|
53
|
+
readonly backend: 'doppler';
|
|
54
|
+
readonly device: unknown | null;
|
|
55
|
+
attachLoraAdapter(adapter: DreamProviderAttachLoRARequest): Promise<{
|
|
56
|
+
adapterId: string;
|
|
57
|
+
rank: number;
|
|
58
|
+
alpha: number;
|
|
59
|
+
scale: number;
|
|
60
|
+
targetModules: string[];
|
|
61
|
+
layerCount: number;
|
|
62
|
+
}>;
|
|
63
|
+
detachLoraAdapter(adapterId?: string | null): Promise<{
|
|
64
|
+
detached: true;
|
|
65
|
+
adapterId: string | null;
|
|
66
|
+
}>;
|
|
67
|
+
generate(request: string | DreamProviderGenerateRequest): Promise<DreamProviderGenerateResult>;
|
|
68
|
+
}
|
|
69
|
+
|
|
70
|
+
export declare function wrapPipelineAsDreamProvider(
|
|
71
|
+
pipeline: {
|
|
72
|
+
generate: (...args: unknown[]) => AsyncIterable<unknown>;
|
|
73
|
+
setLoRAAdapter: (adapter: unknown | null) => void;
|
|
74
|
+
getActiveLoRA: () => unknown | null;
|
|
75
|
+
[key: string]: unknown;
|
|
76
|
+
},
|
|
77
|
+
resolved?: {
|
|
78
|
+
modelId?: string;
|
|
79
|
+
manifest?: Record<string, unknown>;
|
|
80
|
+
device?: unknown | null;
|
|
81
|
+
}
|
|
82
|
+
): DreamCausalLmBaseProvider;
|
|
@@ -1,3 +1,110 @@
|
|
|
1
|
+
import { LORA_MODULE_ALIASES } from '../inference/pipelines/text/lora.js';
|
|
2
|
+
|
|
3
|
+
async function collectText(iterable) {
|
|
4
|
+
let result = '';
|
|
5
|
+
for await (const chunk of iterable) {
|
|
6
|
+
if (typeof chunk === 'string') {
|
|
7
|
+
result += chunk;
|
|
8
|
+
} else if (chunk && typeof chunk.text === 'string') {
|
|
9
|
+
result += chunk.text;
|
|
10
|
+
}
|
|
11
|
+
}
|
|
12
|
+
return result;
|
|
13
|
+
}
|
|
14
|
+
|
|
15
|
+
function getPipelineModelId(pipeline, resolved = {}) {
|
|
16
|
+
return String(
|
|
17
|
+
resolved.modelId ||
|
|
18
|
+
pipeline?.manifest?.meta?.modelId ||
|
|
19
|
+
pipeline?.manifest?.modelId ||
|
|
20
|
+
''
|
|
21
|
+
);
|
|
22
|
+
}
|
|
23
|
+
|
|
24
|
+
function asFloat32Array(value, label) {
|
|
25
|
+
if (value instanceof Float32Array) return value;
|
|
26
|
+
if (Array.isArray(value) || ArrayBuffer.isView(value)) {
|
|
27
|
+
const out = new Float32Array(value);
|
|
28
|
+
if (out.length > 0) return out;
|
|
29
|
+
}
|
|
30
|
+
throw new Error(`wrapPipelineAsDreamProvider: ${label} must be a non-empty numeric array.`);
|
|
31
|
+
}
|
|
32
|
+
|
|
33
|
+
function normalizeTargetModule(moduleName) {
|
|
34
|
+
const key = String(moduleName || '').trim();
|
|
35
|
+
const normalized = LORA_MODULE_ALIASES[key];
|
|
36
|
+
if (!normalized) {
|
|
37
|
+
throw new Error(`wrapPipelineAsDreamProvider: unsupported LoRA target module "${moduleName}".`);
|
|
38
|
+
}
|
|
39
|
+
return normalized;
|
|
40
|
+
}
|
|
41
|
+
|
|
42
|
+
function normalizeDreamLayerMap(layers, rank, alpha) {
|
|
43
|
+
if (!(layers instanceof Map)) {
|
|
44
|
+
throw new Error('wrapPipelineAsDreamProvider: attachLoraAdapter requires layers as a Map.');
|
|
45
|
+
}
|
|
46
|
+
const normalizedLayers = new Map();
|
|
47
|
+
for (const [layerKey, moduleMap] of layers.entries()) {
|
|
48
|
+
const layerIndex = Number(layerKey);
|
|
49
|
+
if (!Number.isInteger(layerIndex) || layerIndex < 0) {
|
|
50
|
+
throw new Error(`wrapPipelineAsDreamProvider: invalid LoRA layer index "${layerKey}".`);
|
|
51
|
+
}
|
|
52
|
+
if (!moduleMap || typeof moduleMap !== 'object' || Array.isArray(moduleMap)) {
|
|
53
|
+
throw new Error(`wrapPipelineAsDreamProvider: layer ${layerIndex} module map must be an object.`);
|
|
54
|
+
}
|
|
55
|
+
const normalizedModules = {};
|
|
56
|
+
for (const [moduleKey, weights] of Object.entries(moduleMap)) {
|
|
57
|
+
const moduleName = normalizeTargetModule(moduleKey);
|
|
58
|
+
const a = asFloat32Array(weights?.a, `layer ${layerIndex} ${moduleName}.a`);
|
|
59
|
+
const b = asFloat32Array(weights?.b, `layer ${layerIndex} ${moduleName}.b`);
|
|
60
|
+
normalizedModules[moduleName] = {
|
|
61
|
+
a,
|
|
62
|
+
b,
|
|
63
|
+
rank,
|
|
64
|
+
alpha,
|
|
65
|
+
scale: rank > 0 ? alpha / rank : 1,
|
|
66
|
+
};
|
|
67
|
+
}
|
|
68
|
+
if (Object.keys(normalizedModules).length === 0) {
|
|
69
|
+
throw new Error(`wrapPipelineAsDreamProvider: layer ${layerIndex} has no LoRA modules.`);
|
|
70
|
+
}
|
|
71
|
+
normalizedLayers.set(layerIndex, normalizedModules);
|
|
72
|
+
}
|
|
73
|
+
if (normalizedLayers.size === 0) {
|
|
74
|
+
throw new Error('wrapPipelineAsDreamProvider: attachLoraAdapter requires at least one layer.');
|
|
75
|
+
}
|
|
76
|
+
return normalizedLayers;
|
|
77
|
+
}
|
|
78
|
+
|
|
79
|
+
function normalizeDreamLoraAdapter(adapter) {
|
|
80
|
+
const adapterId = String(adapter?.adapterId || adapter?.name || '').trim();
|
|
81
|
+
if (!adapterId) {
|
|
82
|
+
throw new Error('wrapPipelineAsDreamProvider: attachLoraAdapter requires adapterId.');
|
|
83
|
+
}
|
|
84
|
+
const rank = Number(adapter?.rank);
|
|
85
|
+
if (!Number.isInteger(rank) || rank <= 0) {
|
|
86
|
+
throw new Error('wrapPipelineAsDreamProvider: attachLoraAdapter requires positive integer rank.');
|
|
87
|
+
}
|
|
88
|
+
const alpha = Number.isFinite(Number(adapter?.alpha))
|
|
89
|
+
? Number(adapter.alpha)
|
|
90
|
+
: Number(adapter?.scale) * rank;
|
|
91
|
+
if (!Number.isFinite(alpha)) {
|
|
92
|
+
throw new Error('wrapPipelineAsDreamProvider: attachLoraAdapter requires alpha or scale.');
|
|
93
|
+
}
|
|
94
|
+
const targetModules = Array.isArray(adapter?.targetModules)
|
|
95
|
+
? adapter.targetModules.map(normalizeTargetModule)
|
|
96
|
+
: [];
|
|
97
|
+
return {
|
|
98
|
+
name: adapterId,
|
|
99
|
+
version: adapter?.version,
|
|
100
|
+
baseModel: adapter?.baseModel || adapter?.baseModelId,
|
|
101
|
+
rank,
|
|
102
|
+
alpha,
|
|
103
|
+
targetModules,
|
|
104
|
+
layers: normalizeDreamLayerMap(adapter.layers, rank, alpha),
|
|
105
|
+
};
|
|
106
|
+
}
|
|
107
|
+
|
|
1
108
|
/**
|
|
2
109
|
* Adapts a raw Doppler pipeline into the model handle shape that
|
|
3
110
|
* createDopplerProvider() expects, without triggering a load.
|
|
@@ -11,24 +118,12 @@ export function wrapPipelineAsHandle(pipeline, resolved = {}) {
|
|
|
11
118
|
throw new Error('wrapPipelineAsHandle requires a loaded pipeline with a generate() method.');
|
|
12
119
|
}
|
|
13
120
|
|
|
14
|
-
async function collectText(iterable) {
|
|
15
|
-
let result = '';
|
|
16
|
-
for await (const chunk of iterable) {
|
|
17
|
-
if (typeof chunk === 'string') {
|
|
18
|
-
result += chunk;
|
|
19
|
-
} else if (chunk && typeof chunk.text === 'string') {
|
|
20
|
-
result += chunk.text;
|
|
21
|
-
}
|
|
22
|
-
}
|
|
23
|
-
return result;
|
|
24
|
-
}
|
|
25
|
-
|
|
26
121
|
return {
|
|
27
122
|
get loaded() {
|
|
28
123
|
return pipeline.isLoaded === true;
|
|
29
124
|
},
|
|
30
125
|
get modelId() {
|
|
31
|
-
return
|
|
126
|
+
return getPipelineModelId(pipeline, resolved);
|
|
32
127
|
},
|
|
33
128
|
get manifest() {
|
|
34
129
|
return pipeline.manifest || resolved.manifest || null;
|
|
@@ -82,3 +177,83 @@ export function wrapPipelineAsHandle(pipeline, resolved = {}) {
|
|
|
82
177
|
},
|
|
83
178
|
};
|
|
84
179
|
}
|
|
180
|
+
|
|
181
|
+
export function wrapPipelineAsDreamProvider(pipeline, resolved = {}) {
|
|
182
|
+
if (!pipeline || typeof pipeline.generate !== 'function') {
|
|
183
|
+
throw new Error('wrapPipelineAsDreamProvider requires a loaded pipeline with a generate() method.');
|
|
184
|
+
}
|
|
185
|
+
if (typeof pipeline.setLoRAAdapter !== 'function') {
|
|
186
|
+
throw new Error('wrapPipelineAsDreamProvider requires pipeline.setLoRAAdapter(adapter).');
|
|
187
|
+
}
|
|
188
|
+
if (typeof pipeline.getActiveLoRA !== 'function') {
|
|
189
|
+
throw new Error('wrapPipelineAsDreamProvider requires pipeline.getActiveLoRA().');
|
|
190
|
+
}
|
|
191
|
+
|
|
192
|
+
const adapters = new Map();
|
|
193
|
+
|
|
194
|
+
return {
|
|
195
|
+
get modelId() {
|
|
196
|
+
return getPipelineModelId(pipeline, resolved);
|
|
197
|
+
},
|
|
198
|
+
get manifest() {
|
|
199
|
+
return pipeline.manifest || resolved.manifest || null;
|
|
200
|
+
},
|
|
201
|
+
get backend() {
|
|
202
|
+
return 'doppler';
|
|
203
|
+
},
|
|
204
|
+
get device() {
|
|
205
|
+
return resolved.device || pipeline.gpuContext?.device || pipeline.device || null;
|
|
206
|
+
},
|
|
207
|
+
async attachLoraAdapter(adapter) {
|
|
208
|
+
const normalized = normalizeDreamLoraAdapter(adapter);
|
|
209
|
+
adapters.set(normalized.name, normalized);
|
|
210
|
+
pipeline.setLoRAAdapter(normalized);
|
|
211
|
+
return {
|
|
212
|
+
adapterId: normalized.name,
|
|
213
|
+
rank: normalized.rank,
|
|
214
|
+
alpha: normalized.alpha,
|
|
215
|
+
scale: normalized.rank > 0 ? normalized.alpha / normalized.rank : 1,
|
|
216
|
+
targetModules: normalized.targetModules,
|
|
217
|
+
layerCount: normalized.layers.size,
|
|
218
|
+
};
|
|
219
|
+
},
|
|
220
|
+
async detachLoraAdapter(adapterId = null) {
|
|
221
|
+
const id = adapterId == null ? null : String(adapterId);
|
|
222
|
+
if (id && !adapters.has(id)) {
|
|
223
|
+
throw new Error(`wrapPipelineAsDreamProvider: cannot detach unknown LoRA adapter "${id}".`);
|
|
224
|
+
}
|
|
225
|
+
if (id) adapters.delete(id);
|
|
226
|
+
if (!id) adapters.clear();
|
|
227
|
+
pipeline.setLoRAAdapter(null);
|
|
228
|
+
return { detached: true, adapterId: id };
|
|
229
|
+
},
|
|
230
|
+
async generate(request = {}) {
|
|
231
|
+
const payload = typeof request === 'string' ? { prompt: request } : request;
|
|
232
|
+
const prompt = payload?.prompt;
|
|
233
|
+
if (prompt == null) {
|
|
234
|
+
throw new Error('wrapPipelineAsDreamProvider.generate requires prompt.');
|
|
235
|
+
}
|
|
236
|
+
const loraAdapterId = payload?.loraAdapterId == null ? null : String(payload.loraAdapterId);
|
|
237
|
+
const samplingOptions = payload?.samplingOptions && typeof payload.samplingOptions === 'object'
|
|
238
|
+
? payload.samplingOptions
|
|
239
|
+
: {};
|
|
240
|
+
const targetAdapter = loraAdapterId ? adapters.get(loraAdapterId) : null;
|
|
241
|
+
if (loraAdapterId && !targetAdapter) {
|
|
242
|
+
throw new Error(`wrapPipelineAsDreamProvider.generate requested unknown LoRA adapter "${loraAdapterId}".`);
|
|
243
|
+
}
|
|
244
|
+
const previousAdapter = pipeline.getActiveLoRA();
|
|
245
|
+
pipeline.setLoRAAdapter(targetAdapter);
|
|
246
|
+
try {
|
|
247
|
+
const text = await collectText(pipeline.generate(prompt, samplingOptions));
|
|
248
|
+
return {
|
|
249
|
+
text,
|
|
250
|
+
useLora: Boolean(targetAdapter),
|
|
251
|
+
baseModelId: getPipelineModelId(pipeline, resolved),
|
|
252
|
+
loraAdapterId,
|
|
253
|
+
};
|
|
254
|
+
} finally {
|
|
255
|
+
pipeline.setLoRAAdapter(previousAdapter || null);
|
|
256
|
+
}
|
|
257
|
+
},
|
|
258
|
+
};
|
|
259
|
+
}
|
|
@@ -30,11 +30,16 @@ Checked-in config inventory:
|
|
|
30
30
|
| `gemma3/gemma-3-1b-it-f16-af32.json` | `gemma-3-1b-it-f16-af32` | F16 weights, F32 activation policy |
|
|
31
31
|
| `gemma3/gemma-3-1b-it-q4k-ehf16-af32.json` | `gemma-3-1b-it-q4k-ehf16-af32` | Q4K row layout with execution-v1 graph |
|
|
32
32
|
| `gemma3/gemma-3-270m-it-q4k-ehf16-af32.json` | `gemma-3-270m-it-q4k-ehf16-af32` | Q4K row layout with execution-v1 graph |
|
|
33
|
+
| `gemma3/gemma-3-270m-it-q4k-ehf16-af16.json` | `gemma-3-270m-it-q4k-ehf16-af16` | Experimental Q4K f16 activation repro; blocked by layer-8 prefill attention-core NaNs on local verify |
|
|
33
34
|
| `gemma3/translategemma-4b-1b-enes-q4k-ehf16-af32.json` | `translategemma-4b-1b-enes-q4k-ehf16-af32` | TranslateGemma distill/variant artifact |
|
|
34
35
|
| `gemma3/translategemma-4b-it-q4k-ehf16-af32.json` | `translategemma-4b-it-q4k-ehf16-af32` | Text-only TranslateGemma artifact |
|
|
35
36
|
| `gemma4/gemma-4-e2b-it-q4k-ehf16-af32.json` | `gemma-4-e2b-it-q4k-ehf16-af32` | Gemma 4 E2B Q4K, execution-v1 graph |
|
|
36
37
|
| `gemma4/gemma-4-e2b-it-q4k-ehf16-af32-int4ple.json` | `gemma-4-e2b-it-q4k-ehf16-af32-int4ple` | Gemma 4 E2B Q4K with INT4 PLE policy |
|
|
38
|
+
| `gemma4/gemma-4-12b-it-text-q4k-ehf16-af32.json` | `gemma-4-12b-it-text-q4k-ehf16-af32` | Gemma 4 12B text-only Q4K, execution-v1 graph |
|
|
39
|
+
| `gemma4/gemma-4-12b-it-text-q4k-ehf16-af16.json` | `gemma-4-12b-it-text-q4k-ehf16-af16` | Gemma 4 12B text-only Q4K f16 activation sibling over the existing weight pack |
|
|
40
|
+
| `gemma4/gemma-4-12b-it-text-w4a16-ct-ehf16-af16.json` | `gemma-4-12b-it-text-w4a16-ct-ehf16-af16` | Gemma 4 12B official W4A16 compressed-tensors QAT reference-dequant lane |
|
|
37
41
|
| `gemma4/gemma-4-moe-q4k-ehf16-af32.json` | `gemma-4-moe-q4k-ehf16-af32` | Gemma 4 MoE Q4K |
|
|
42
|
+
| `diffusiongemma/diffusiongemma-26b-a4b-it-q4k-ehf16-af16.json` | `diffusiongemma-26b-a4b-it-q4k-ehf16-af16` | DiffusionGemma text-only Q4K f16 activation block-diffusion artifact |
|
|
38
43
|
| `gpt-oss-20b-f16-xmxfp4.json` | `gpt-oss-20b-f16-xmxfp4` | GPT-OSS F16/XMXFP4 config |
|
|
39
44
|
| `janus/janus-pro-1b-text-q4k-ehaf16.json` | `janus-pro-1b-text-q4k-ehaf16` | Janus text-only Q4K artifact |
|
|
40
45
|
| `lfm2/lfm2.5-1.2b-instruct-q4k-ehf16-af32.json` | `lfm2.5-1.2b-instruct-q4k-ehf16-af32` | LFM2 hybrid conv/attention scheduling |
|