[wangyao-i]

Purpose

Add an Ascend NPU AFD role split for DeepSeek-V4.1 Flash (DeepseekV41ForCausalLM) on pinned vLLM 5df63a7 (0.32.0) and vLLM-Ascend bd42eea0.

Current validated candidate: 42a893ec46454d398388fb13e5b9add125f3494d, rebased onto main 127d8ec.

Scope and implementation

• Attention retains the native DSA/HC/Engram decoder forward, candidate buffers and shared attention state. Its MoE slot becomes an AFD remote proxy; async CAM keeps routing and shared experts on Attention.
• FFN constructs native DeepseekV41MoE and exposes routed-expert weights through get_async_cam_layered_weights() to the async runner. Synchronous gate-on-FFN execution retains compute_ffn_output(); the obsolete legacy gate-on-Attention helper is removed.
• Engram tables stay on Attention. FFN declares has_engram=False and excludes Engram checkpoint tensors. Decoder replay is disabled on Attention.
• Weight classification accepts an optional model. prefix before excluding vision/aligner/image/MTP tensors; original names passed to the native loader remain unchanged.
• Register the NPU-only AFD architecture and reject non-NPU model-config resolution.
• Reuse V4 routing, CAM exchange, shared-expert layout and expert-weight extraction, with the V4.1 already_sequence_parallel forward argument.

Test plan and results

Local checks

• Focused model/registration tests plus async layered/runtime selection: 99 passed, 1 skipped (vLLM unavailable locally).
• Ruff check / format check: all seven changed Python files pass.
• Role-filter coverage is now at tests/unit/model_executor/models/npu/test_deepseek_v41.py. It reads real production constants rather than duplicating them, covers six prefixed non-text regressions, preserves original loader names, and checks rejection of legacy async FFN execution.
• Architecture rewrite/NPU-only guard coverage is in tests/unit/model_executor/models/test_model_utils.py; package tests exercise the external registration seam.
• An earlier broader Windows run had an unrelated source-distribution manifest test failure; this is not a claim that the full local unit suite passes.

Ascend server validation — reported for clean head 42a893e

Server validation record, sections 10.4–10.6.

Item	Reported configuration/result
Runtime	vLLM source build 0.32.0+empty at 5df63a7; Ascend bd42eea0; torch_npu 2.10.0.post7
Hardware / weights	A3, real DeepSeek-V4.1-Flash W8A8 checkpoint
Topology	Attention 1 die, FFN 12 dies (DP12); async CAM, eager
Unit tests	Server report states the full tests/unit suite passed; numeric totals are not provided
Startup / generation	Both roles started; temperature-0 smoke returned correct answers for 1+1 and 2+2
Weight classification	143,322 checkpoint names across five ownership configurations: no unclassified-path exceptions

The report identifies a clean checkout without the earlier server-local prefix patch. Exact current executor selection and per-run raw artifacts have not been independently inspected here; the old non-layered launch settings in section 9 are historical.

Reported benchmark values (vllm bench serve, OpenAI chat, random inputs, --ignore-eos, concurrency 4):

Input/output tokens	Requests succeeded/failed	Output tok/s	Mean TPOT	Median TTFT
256 / 64	32 / 0	38.14	90.79 ms	366.70 ms
512 / 128	32 / 0	41.24	91.59 ms	536.79 ms

These values are reproduced from the server report. There is no native baseline or repeated-run statistical comparison, so no speedup or performance-regression conclusion is claimed.

GSM8K — bounded server evaluation

• 194/200 correct (97.00%), concurrency 4, temperature 0, max_tokens=3072.
• Uses a custom stdlib evaluator on GSM8K test data, with a step-by-step prompt and numeric answer extraction. This is not an official lm-eval-harness/GSM8K-7 result; evaluator equivalence has not been established.
• The 1,319-question evaluation was interrupted for resource sharing and remains incomplete.
• No matched native-versus-AFD accuracy baseline was supplied, so individual answer errors are not attributed to or ruled out as AFD effects.

Restrictions and remaining validation

Construction rejects non-NPU execution, multimodal execution without --language-model-only, async CAM without compute_gate_on_attention=true, sequence-parallel MoE without Attention-side gating, EPLB, PP/PCP/DCP sizes other than 1, speculative decoding, LoRA, SWA bounded replay, and Engram with DBO/microbatching.

• Hardware evidence is limited to single-node A1/F12 W8A8 eager serving. Graph, TP/SP variants, synchronous CAMP2P, W4A8, PD, multi-node and long-duration stability are not qualified by this record.
• One vector core exception (507035) caused an EngineCore crash during validation. Restart recovered service; its root cause remains undetermined.
• The server report flags one stale test link in docs/design/module/model_integration.md; it still needs to point to tests/unit/model_executor/models/npu/test_deepseek_v41.py.

Docs impact

docs/design/module/model_integration.md documents role ownership, checkpoint filtering and support restrictions. Its hardware-validation wording predates the latest server report; this PR description and the linked record provide the current bounded results.

[jiangkuaixue123]

Found a checkpoint-filtering failure in the text-only role split. Reviewed against pinned vLLM-Ascend bd42eea0; NPU construction and execution were not run.

[jiangkuaixue123]

[P1] Normalize the optional model prefix before filtering non-text weights

The pinned Ascend implementation explicitly supports model.image_end in its text-only checkpoint path (see test_v41_text_only_load_skips_checkpoint_vision_weights). This check only drops unprefixed image_*, vision.*, and aligner.* names. For model.image_end, _weight_layer_path() returns None, so the tensor is sent to both AFD roles. Because this adapter subclasses the text LLM directly, it bypasses the native multimodal wrapper's prefix normalization, and the inherited _load_model_weights() reaches params_dict['model.image_end'], raising KeyError since neither role allocates that parameter. Normalize an optional model. prefix for non-text/deferred-weight classification, and cover prefixed checkpoint names as well as the unprefixed ones.