fix(npu): disable shared TP for Async CAM model SP

[jiaran-king]

Observed problem and introduction

NPU Async CAM Attention can run with model TP and SP enabled while its standalone shared MLP still holds TP-sharded weights. SP assigns different token rows to the TP ranks, but the shared MLP's native row-parallel AllReduce assumes those ranks are computing partial results for the same tokens. It therefore sums contributions from different tokens and produces incorrect shared-expert output. This can corrupt generation and accuracy without causing a shape error or process failure.

The introduction chain is:

• AFD #363 placed the standalone shared MLP on Attention for Async CAM. With the earlier Ascend implementation, its is_sequence_parallel=True / disable_tp setting selected complete shared weights.
• Ascend #13946, commit 92616f3a, changed shared linear layers to select replication through enable_shared_expert_dp; shared layers no longer honor the generic disable_tp flag.
• AFD #407 upgraded to Ascend 8d4409d6 without adapting this standalone shared path. #425 subsequently integrated the upgrade into main.

This is an AFD compatibility gap exposed by the Ascend dependency upgrade. Both MRV1 and MRV2 reuse the affected shared branch.

Affected models and configurations

The incorrect layout requires all of the following:

• NPU CAMAsyncAFDConnector, on the Attention role, with compute_gate_on_attention=true.
• A model with standalone shared experts using the DeepSeek-V2-family adapter: model types deepseek, deepseek_v2, deepseek_v3, deepseek_v32, or glm_moe_dsa. Existing MRV2 admission remains limited to V2/V3 MLA within this family.
• Effective model MoE SP, with Attention TP > 1, EP enabled and PP = 1. The fix checks the effective SP decision after native FlashComm derivation; requesting FlashComm alone is insufficient. The hardware test used A DP2/TP2, enable_flashcomm1=true, and allgather_reducescatter.
• Shared-expert DP unset/false, leaving the shared projections TP-sharded.

TP alone is valid when the shared inputs contain the same tokens on each rank. TP=1 and already-replicated shared weights do not have this mismatch. DeepSeek-V4 uses a separate implementation and is outside this fix.

Fix and behavior scope

For the affected TP + SP path, disable only shared-expert TP by replicating complete shared weights on each Attention rank. Keep the model's other TP/SP behavior.

After native FlashComm/SP derivation and before model weight allocation, select Ascend's existing enable_shared_expert_dp=true mechanism for the affected Attention path. This makes both shared projections use tp_size=1: each rank computes the complete shared MLP for its own SP token rows, without a shared TP AllReduce. Despite the flag's name, this does not change the configured model DP or TP sizes. The existing native placement check is retained.

Configuration	Result after this change
Affected Attention path, effective SP on, shared DP unset/false	Shared weights become complete on each rank; shared TP becomes 1. Model TP/SP remain enabled.
Effective model SP off, including TP=1 or this family's PP > 1 path	No automatic shared setting override; preserve the requested shared layout.
Shared DP already true in a configuration accepted by the existing validator	Preserve the replicated shared layout.
FFN role, other connectors, gate-on-FFN, or non-AFD	No automatic override from this fix.
DeepSeek-V4	Existing implementation and MRV2 shared-DP restriction remain.

MRV2 configuration admission is unchanged. Its existing rejection of explicit enable_shared_expert_dp=true remains for V2/V3 and V4, together with the finegrained-TP restriction. This PR changes the effective Ascend shared-weight layout only in the affected SP path; the source additional_config mapping is preserved. It does not introduce a new admission rule or expand the accepted parameter set.

The production diff is +26/-1 lines in one file, afd_plugin/compat/patches/npu/ascend_config.py. The validator, its existing runtime tests, the shared MLP constructor and forward helper match main. Existing deployment documentation and focused factory scope coverage are updated. Replication increases shared-weight memory per affected Attention rank; performance and memory equivalence are not claimed.

Validation and evidence limits

Final PR commit: f44bb989cb10eb4217435b575ce2da94d6cfceb2. The shared-layout factory code is identical to the previously tested commit f6bc4961fe2f8f917f9505b9ed2769e21f424d90; the MRV2 admission change has been removed. The retained MRV1 production path is unchanged by that removal.

Current closeout checks:

• Factory/lifecycle tests: 5 passed, 1 skipped locally. The skipped existing splitting-ops test requires the pinned vLLM source via AFD_TEST_VLLM_SOURCE. The effective-SP scope test covers roles, connectors, gate placement, model family, SP off, PP > 1, explicit shared settings and preservation of the input configuration.
• Repository pre-commit checks passed, including Ruff, mypy 3.10 and Markdown checks.
• The validator and its runtime test file are byte-identical to main. Evaluating the restored MRV2 admission block against the recorded MRV2 launch options confirms that its explicit shared-DP=true request is rejected.

Reusable MRV1 hardware evidence from f6bc4961:

• DeepSeek-V2-Lite, Ascend 910C, CANN 9.1.0, torch 2.10.0, torch-npu 2.10.0.post4, vLLM ced6857a (0.30.0), Ascend 8d4409d6; Async CAM eager, A DP2/TP2 and F DP4/TP1 with EP. MRV1 explicitly requested shared DP=false, exercising the automatic layout correction. Both the Python extension and CAM operator package were compiled for that hardware run; the upstream source trees were clean.
• During real requests, all four Attention ranks retained model TP=2, SP=true; both shared projections used tp_size=1 with complete shared weights (gate_up_proj shape [5632, 2048]). This confirms SP remained active and the shared branch actually used the intended layout.
• MRV1 GSM8K: 41/128 (32.03%), compared with the existing native reference's 39/128 (30.47%) on the same 128 questions. The raw samples were rechecked: all prompt and target hashes match. Accuracy uses flexible answer extraction; identical generated text is not required.

MRV2 hardware closeout on f44bb98 (2026-10-13):

• DeepSeek-V2-Lite on Ascend 910C with the same pinned vLLM/Ascend commits and CANN/torch/torch-npu versions listed above. All 528 tracked source files matched the final PR commit; both upstream source trees were clean. The matching, previously compiled extension and CAM operators were reused, with their loaded hashes verified.
• Async CAM eager, A DP2/TP2 and F DP4/TP1 with EP. Both roles explicitly requested enable_shared_expert_dp=false, which the unchanged MRV2 validator accepts. Attention enabled FlashComm1 with allgather_reducescatter, exercising the automatic shared-layout correction.
• During real requests the actual runner was AFDNPUAttentionModelRunnerV2. All four Attention ranks retained model TP=2, SP=true; effective shared DP became true, both shared projections used tp_size=1, and each rank held complete shared weights (gate_up_proj shape [5632, 2048]). All 70 live shared-branch observations had positive token rows, covering every Attention rank within the request window.
• 4/4 serial and 12/12 concurrent requests returned correct numeric answers, with concurrency 12 and prompt lengths 57/121/186/250 tokens. Four elementary arithmetic questions used independently checked answers 12/27/6/42 and flexible numeric extraction. This is a functional smoke test, not a GSM8K accuracy result.
• The launcher exited with code 0; all model processes were cleared from the devices and the test task was stopped with storage preserved. No production code or PR commit changed for this verification.

Correction to the earlier smoke-test claim: both historical 25-request accounting runs completed, including concurrency 12 and prompt token counts 305/369/433/497. However, their checker accepted B / 62.50% / 2.17, while the question gives 16/31 = 51.61% and 19/7 = 2.71, with no correct supplied option. These runs provide request-path/concurrency smoke evidence only, not mathematical correctness or accuracy evidence. The independent GSM8K result above is the accuracy evidence.

The historical MRV2 command explicitly requested shared DP=true and remains rejected by the unchanged validator; that run is not used to qualify this PR. The new accepted-configuration smoke run above verifies MRV2 startup and request execution for the tested topology. These V2-Lite eager runs do not qualify other models, graph execution or performance. No GSM8K rerun was performed for the MRV2 closeout.