[model, ci] feat: migrate deepseek_v3 to transformers v5

[TimYangst]

What does this PR do?

Migrates deepseek_v3 from the v4 runtime monkey-patch path to the
transformers v5 patchgen + self-contained generated modeling path
(Pattern B: v4↔v5 coexist, MoE, GPU + NPU). v5 gate is 5.2.0.

Checklist Before Starting

• PR title follows [{modules}] {type}: {description} format

Test

Ran against transformers==5.2.0 on an 8×GPU box.

• python -m veomni.patchgen.check_patchgen — all generated files up to date.

• make quality — clean.

• pytest tests/models/test_models_patch.py -k deepseek_v3 — PASSED (1 passed in 20.58s).
  Validates HF↔VeOmni fwd/bwd parity for the patched v5 model.

• pytest tests/e2e/test_e2e_parallel.py -k deepseek_v3_v5 — PASSED
  (1 passed in 168.89s). Compares 4 parallel configs under SP/EP.

  grad_norm
    run 1: 1.37599146, 1.25248539
    run 2: 1.37604284, 1.25065529
    run 3: 1.37589073, 1.25190842
    run 4: 1.37582600, 1.25116014
  loss
    run 1: 12.17874718, 11.68118858
    run 2: 12.17911124, 11.68350935
    run 3: 12.17874718, 11.68191600
    run 4: 12.17911100, 11.68174267
  

• pytest tests/distributed/test_fsdp_equivalence.py -k deepseek_v3_v5 —
  PASSED (1 passed in 84.21s). Single-GPU (no FSDP) vs 2-GPU FSDP2.

  grad_norm
    single_gpu: 1.34375000, 1.25781250
    fsdp2_2gpu: 1.34375000, 1.25781250        # exact match
  loss
    single_gpu: 12.16680050, 11.83021641
    fsdp2_2gpu: 12.18102837, 11.84898996      # diff expected, see test docstring
  

API and Usage Example

N/A — internal model registration; public API unchanged.

Design & Code Changes

• Added deepseek_v3_gpu_patch_gen_config.py and
  deepseek_v3_npu_patch_gen_config.py with patches for
  DeepseekV3NaiveMoe (fused gate_up_proj/down_proj layout, drops
  upstream @use_experts_implementation, eager + fused branches),
  DeepseekV3TopkRouter.forward (autocast-disabled fp32 router for
  actor/rollout parity), DeepseekV3ForCausalLM.forward (fused-CE path)
  and get_parallel_plan.
• Liger kernels and deterministic Triton RoPE + batch-invariant RMSNorm
  are kept out of the generated file and re-applied at runtime in
  __init__.py based on VEOMNI_USE_LIGER_KERNEL, mirroring v4 behavior.
  LigerSwiGLUMLP is intentionally skipped — shared_experts passes an
  intermediate_size kwarg that Liger's MLP does not accept.
• Added checkpoint_tensor_converter.py for HF per-expert →
  v5 fused runtime conversion; registered as staticmethod on all v5
  model classes.
• Wired __init__.py with Pattern B: v5 gate at 5.2.0 with GPU/NPU
  branching; falls back to the v4 monkey-patch path below that.
• Generated patched_modeling_deepseek_v3_{gpu,npu}.py committed.
• Added deepseek_v3 entry to _TEST_CASES_TRANSFORMERS_V5 and a
  deepseek_v3_v5 param to test_e2e_parallel.py.
• Updated sync_weight_deepseek_v3 to skip per-expert stacking when HF
  state dict already uses the v5 fused layout.

Checklist Before Submitting

• Read the Contribute Guide
• Applied pre-commit checks
• Added/updated documentation
• Added tests to CI workflow (or explained why not feasible)

[gemini-code-assist]

Code Review

This pull request introduces comprehensive support for DeepseekV3 models, specifically targeting transformers library versions 5.2.0 and above. Key changes include a new runtime checkpoint tensor converter that automatically handles the conversion of HuggingFace's per-expert checkpoint format to a fused v5 format, thereby streamlining the loading process by removing the need for offline merging. The DeepseekV3 modeling code has been refactored to dynamically apply GPU or NPU-specific patches and optimized kernels (such as Liger kernels or batch-invariant alternatives) based on the execution environment and transformers version. These patches enhance the fused Mixture-of-Experts (MoE) implementation, ensure numerical parity for router autocast behavior, and enable a fused cross-entropy path in the causal language model. Corresponding e2e and model patching tests have been added to validate the DeepseekV3 integration. No specific feedback was provided in the review comments.