Fix pure-text AFD model config identity

[ShwStone]

Purpose

Fix the model configuration regression introduced by #144 and reported in
#172.

get_afd_model_config() copied and replaced hf_config without updating
hf_text_config. For pure-text models, this broke the original identity:

model_config.hf_text_config is model_config.hf_config

vLLM-Ascend consequently treated DeepSeekV2-Lite as a model with a distinct
nested text configuration. Under Attention TP2 with FlashComm1/SP, MLA skipped
the required query/KV gather and passed TP-local x together with full-batch
RoPE cos/sin, causing npu_interleave_rope to fail.

The regression was confirmed across #144:

• 220c7626de5195bdcf24271659f0c1519a71f2a3: passes
• 355a1cda4a1f170b0fd5db0029eba8c521151241: fails

Issue

• Related issue(s): [Bug]: PR #144 breaks pure-text model config identity and causes Async CAM TP2 FlashComm1 RoPE failure #172
• Closing keyword, only if fully resolved: Closes [Bug]: PR #144 breaks pure-text model config identity and causes Async CAM TP2 FlashComm1 RoPE failure #172

Scope

• In scope:
  • Preserve hf_config/hf_text_config identity when converting a pure-text
    model config to its AFD architecture.
  • Preserve a genuinely distinct nested hf_text_config unchanged.
  • Add unit coverage for both configuration layouts.
• Out of scope:
  • Changes to vLLM or vLLM-Ascend.
  • Changes to FlashComm1/SP scheduling or MLA implementation.
  • Changes to Async CAM communication or MoE ubatching.
  • Changes to multimodal model configuration semantics.

Implementation Notes

get_afd_model_config() now creates one private hf_config copy, updates its
architecture, and assigns it to the copied ModelConfig.

When the source is a pure-text configuration:

model_config.hf_text_config is model_config.hf_config

the copied model config points both fields to the new private hf_config.
When the source has a genuinely distinct nested hf_text_config, that object
remains distinct and unchanged.

This is a plugin-owned model configuration fix. It does not patch or modify
vLLM/vLLM-Ascend source code.

Test Plan

Focused unit coverage:

python -m pytest -vv \
  tests/unit/package/test_package.py \
  -k "afd_model_config"

Ascend NPU Async CAM regression coverage:

ASCEND_RT_VISIBLE_DEVICES=0,1,2,3 \
AFD_NPU_E2E_MODEL=/path/to/DeepSeek-V2-Lite \
AFD_NPU_ASYNC_CAM_E2E_DEVICES=0,1,2,3 \
AFD_NPU_E2E_VLLM_BIN=vllm \
AFD_NPU_E2E_STARTUP_TIMEOUT=900 \
python -m pytest -vv -s \
  tests/e2e/models/deepseek_v2_lite/test_async_cam_npu.py::test_deepseek_v2_lite_async_cam_attn_dp1tp2_ffn_dp2ep2_smoke

Test Result

Both targeted tests passed:

• Focused afd_model_config unit tests: passed
• Async CAM Attention DP1TP2 + FFN DP2EP2 NPU E2E smoke test: passed

The E2E result confirms that the existing main-branch Async CAM test can run
with Attention TP2 and FlashComm1 without the previous RoPE batch-dimension
failure.

Docs Impact

• Files updated: none
• If none, reason: this restores the original pure-text ModelConfig invariant
  and does not introduce a new public configuration or runtime option. The
  regression and reproduction are documented in [Bug]: PR #144 breaks pure-text model config identity and causes Async CAM TP2 FlashComm1 RoPE failure #172.

---

Essential PR Checklist

• Purpose is clear and linked to public context when possible.
• Scope is bounded.
• Compatibility with vLLM v0.19.1 is considered.
• No changes are made to the vLLM source checkout.
• Plugin-owned classes or explicit dotted class paths are preferred over monkey patches.
• Any compat shim or monkey patch is isolated, idempotent, version-guarded, documented, and tested.
• Imports remain CPU-safe; CUDA-heavy work is delayed or GPU-gated.
• Validation evidence is included, including skipped GPU tests when applicable.
• Documentation impact is stated.

[Copilot]

Pull request overview

This PR fixes a regression in get_afd_model_config() where converting a pure-text ModelConfig to an AFD architecture broke the invariant hf_text_config is hf_config, which downstream vLLM-Ascend logic relies on to distinguish pure-text vs nested-text/multimodal configurations.

Changes:

• Update get_afd_model_config() to create a single private hf_config copy and (when applicable) re-point hf_text_config to preserve pure-text object identity.
• Add unit tests covering both layouts: pure-text identity preservation and nested hf_text_config preservation.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated no comments.

File	Description
afd_plugin/model_executor/models/model_utils.py	Preserves hf_config/hf_text_config identity for pure-text configs while keeping truly distinct nested text configs unchanged.
tests/unit/package/test_package.py	Adds targeted unit coverage for both configuration layouts to prevent regressions.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[jiangkuaixue123]

Fix DCO

[jiangkuaixue123]

Would it be safer to use deepcopy(model_config) here and then update afd_model_config.hf_config.architectures? deepcopy preserves aliasing within the copied object graph, so the pure-text invariant hf_text_config is hf_config would be retained automatically, and it would reduce the risk of missing other identity relationships if ModelConfig evolves.

The semantic difference is that a genuinely distinct/nested hf_text_config would also become a private copy, whereas this patch intentionally keeps it shared with the source config. Is that sharing required? If not, deepcopy may provide stronger isolation and a simpler implementation. Otherwise, could we document why only hf_config should be copied so this choice is explicit?

[ShwStone]

I confused GPG signing with DCO; I always thought signing was enough to pass DCO. I've already recommitted.

[ShwStone]

@jiangkuaixue123 Agreed — the deepcopy approach is cleaner and preserves the aliasing invariants automatically. I've switched the implementation to deepcopy(model_config) in 38df180 and validated it on real hardware (DeepSeekV2-Lite, Attention TP2 with FlashComm1/SP): the npu_interleave_rope failure is gone and the model runs correctly. Tests updated accordingly. PTAL.