源码证据索引

读取本地三个固定 commit 自动生成。每个摘录有真实行号；源码升级后须重新审计，构建脚本会拒绝版本漂移。

主题	仓库 / 文件	行号
nano-qkv	nano-vllm `nanovllm/models/qwen3.py`	14–91
nano-mlp	nano-vllm `nanovllm/models/qwen3.py`	91–122
nano-decoder	nano-vllm `nanovllm/models/qwen3.py`	120–169
nano-lm-head	nano-vllm `nanovllm/models/qwen3.py`	186–216
nano-embedding	nano-vllm `nanovllm/layers/embed_head.py`	9–66
nano-rope	nano-vllm `nanovllm/layers/rotary_embedding.py`	6–50
nano-rmsnorm	nano-vllm `nanovllm/layers/layernorm.py`	5–50
nano-silu	nano-vllm `nanovllm/layers/activation.py`	6–11
nano-attention	nano-vllm `nanovllm/layers/attention.py`	43–75
nano-store-kv	nano-vllm `nanovllm/layers/attention.py`	11–40
nano-cache-allocation	nano-vllm `nanovllm/engine/model_runner.py`	103–122
nano-prefill-inputs	nano-vllm `nanovllm/engine/model_runner.py`	129–170
nano-decode-inputs	nano-vllm `nanovllm/engine/model_runner.py`	172–188
nano-graph-replay	nano-vllm `nanovllm/engine/model_runner.py`	196–220
nano-graph-capture	nano-vllm `nanovllm/engine/model_runner.py`	223–257
nano-sampler	nano-vllm `nanovllm/layers/sampler.py`	5–12
nano-scheduler	nano-vllm `nanovllm/engine/scheduler.py`	25–92
nano-sequence	nano-vllm `nanovllm/engine/sequence.py`	14–79
nano-blocks	nano-vllm `nanovllm/engine/block_manager.py`	26–120
nano-engine-step	nano-vllm `nanovllm/engine/llm_engine.py`	49–90
vllm-engine-step	vllm `vllm/v1/engine/core.py`	710–740
vllm-request	vllm `vllm/v1/request.py`	165–186
vllm-budget	vllm `vllm/v1/core/sched/scheduler.py`	800–871
vllm-allocation	vllm `vllm/v1/core/sched/scheduler.py`	922–1028
vllm-advance	vllm `vllm/v1/core/sched/scheduler.py`	1888–1907
vllm-output	vllm `vllm/v1/core/sched/scheduler.py`	2349–2458
vllm-contract	vllm `vllm/v1/core/sched/output.py`	245–279
vllm-prefix-lookup	vllm `vllm/v1/core/kv_cache_manager.py`	266–328
vllm-hash	vllm `vllm/v1/core/kv_cache_utils.py`	640–740
vllm-block-pool	vllm `vllm/v1/core/block_pool.py`	681–830
vllm-pack	vllm `vllm/v1/worker/gpu_model_runner.py`	1923–2064
vllm-metadata	vllm `vllm/v1/worker/gpu_model_runner.py`	2102–2196
vllm-slot	vllm `vllm/v1/worker/block_table.py`	348–376
vllm-forward	vllm `vllm/v1/worker/gpu_model_runner.py`	4136–4190
vllm-cudagraph-doc	vllm `docs/design/cuda_graphs.md`	40–169
deepinfer-scope	deepinfer `README.md`	3–52
deepinfer-add	deepinfer `deepinfer/engine/llm_engine.py`	591–631
deepinfer-schedule-type	deepinfer `csrc/engine/scheduler.hpp`	275–343
deepinfer-prefill	deepinfer `csrc/engine/scheduler.hpp`	450–549
deepinfer-decode	deepinfer `csrc/engine/scheduler.hpp`	556–659
deepinfer-metadata	deepinfer `csrc/engine/scheduler.hpp`	667–734
deepinfer-worker	deepinfer `deepinfer/worker/worker.py`	89–188
deepinfer-moe-gate	deepinfer `deepinfer/model_executor/models/v4_modules/mlp.py`	379–401
deepinfer-moe-communication	deepinfer `deepinfer/model_executor/models/v4_modules/mlp.py`	461–511
deepinfer-moe-forward	deepinfer `deepinfer/model_executor/models/v4_modules/mlp.py`	565–616
deepinfer-eplb-policy	deepinfer `deepinfer/model_executor/eplb/policy.py`	12–140
deepinfer-eplb-trigger	deepinfer `deepinfer/model_executor/eplb/__init__.py`	146–165
deepinfer-eplb-migration	deepinfer `deepinfer/model_executor/eplb/__init__.py`	195–278
deepinfer-confidence	deepinfer `deepinfer/model_executor/models/mtp/dflash.py`	46–79
deepinfer-mtp-schedule	deepinfer `csrc/engine/mtp_varlen.hpp`	139–266
deepinfer-accept-rates	deepinfer `csrc/engine/scheduler.hpp`	1239–1250
deepinfer-commit	deepinfer `csrc/engine/scheduler.hpp`	899–1003
deepinfer-cache-counter	deepinfer `csrc/engine/sequence.hpp`	210–221
deepinfer-swa	deepinfer `doc/inference_deployment_qa.md`	235–245
deepinfer-swa-config	deepinfer `deepinfer/environ.py`	194–204
deepinfer-seed	deepinfer `deepinfer/sampling_params.py`	85–93