[Feature] Support apply_liger_kernel_to_deepseek_v3

[Yulong-Cauli]

Summary

Adds Liger Kernel support for DeepSeek-V3 (deepseek_v3 model type), closing the gap left open by #623.

• model/deepseek_v3.py: new lce_forward. In training with labels, skip_logits defaults to True, so logits are never materialized (FLCE path via LigerForCausalLMLoss); otherwise logits are materialized and the HF-native self.loss_function is used. Returns LigerCausalLMOutputWithPast with token_accuracy / predicted_tokens.
• monkey_patch.py: apply_liger_kernel_to_deepseek_v3 with class-level and instance-level patching, registered in MODEL_TYPE_TO_APPLY_LIGER_FN (AutoLigerKernelForCausalLM picks it up automatically).
  • RMSNorm: DeepseekV3RMSNorm → LigerRMSNorm, including attention q_a_layernorm / kv_a_layernorm.
  • SwiGLU: dense MLPs and shared experts (DeepseekV3MLP) → LigerQwen3MoeSwiGLUMLP on all supported versions. Routed experts: batched DeepseekV3Experts → fused LigerExperts on transformers v5+; per-expert patch on transformers v4.
  • RoPE: unsupported — DeepSeek-V3 uses interleaved partial RoPE, incompatible with liger_rotary_pos_emb. rope=True emits a warning_once and skips (no silent behavior change); RoPE is intentionally not listed in the README table.
• Tests: instance monkey-patch test (asserts both v5 and v4 branch outcomes), 4 convergence tests (bf16/fp32 × FLCE/logits) driven by a mini deepseek_v3 config, revert_liger_kernel_to_deepseek_v3 util, README entry.

Details

• The transformers v4 branch (per-expert patching) mirrors the existing mixtral / qwen3_moe patterns but could not be exercised locally (only transformers 5.16.1 is installed here); it is covered by the CI v4 matrix.
• On transformers v5+, LigerExperts accepts DeepseekV3Config (num_local_experts aliases n_routed_experts), and its batched expert weights ((N, 2*I, H) / (N, H, I)) match the HF layout, so pretrained checkpoints load unchanged.

Testing Done

Check	Status
Import	PASS
Instance Patching	PASS
Convergence bf16 FLCE	PASS
Convergence bf16 with_logits	PASS
Convergence fp32 FLCE	PASS
Convergence fp32 with_logits	PASS
Convergence multimodal	SKIP (text-only model)
Checkstyle	PASS

Exact commands run on this change:

• pytest test/transformers/test_monkey_patch.py -k deepseek_v3 → 1 passed

• pytest test/convergence/{bf16,fp32}/test_mini_models.py test/convergence/{bf16,fp32}/test_mini_models_with_logits.py -k deepseek_v3 → 4 passed

• ruff check + ruff format --check clean on all changed files

• Hardware Type: RTX 4060 Laptop GPU (8GB) — transformers 5.16.1, torch 2.11.0+cu128

• run make test to ensure correctness

• run make checkstyle to ensure code style

• run make test-convergence to ensure convergence

Closes #623

[Yulong-Cauli]

@arde171 @kolehma8 @BYHsu — friendly ping for a review when you get a chance.

The branch is now a single commit rebased onto latest main (95b01e9); no conflicts, head is 31c38e6.

Recap (closes #623): adds apply_liger_kernel_to_deepseek_v3 —

• lce_forward built on FusedLinearCrossEntropy, so logits are never materialized during training
• RMSNorm + SwiGLU patching for the dense MLP and shared experts; on transformers v5+ the MoE routed experts are patched via LigerExperts (per-expert patching on v4)
• class-level and instance-level patching, MODEL_TYPE registration
• interleaved partial RoPE is intentionally left unpatched (warns + skips)

Validation:

• Kaggle Tesla T4 (SM 7.5, torch 2.10.0+cu128, transformers 5.0.0): deepseek_v3 3/3 passed — monkey-patch instance test, fp32 FLCE convergence, fp32 logits convergence. (The deepseek_v4 sibling cases skip on that env because transformers 5.0.0 doesn't ship the deepseek_v4 module yet.)
• RTX 4060 Laptop (SM 8.9, transformers 5.16.1): 10/10 passed — deepseek_v3 + deepseek_v4 across bf16/fp32, FLCE and logits paths; the transformers v5 fused-MoE (LigerExperts) path is exercised in both environments.
• Ruff checks are clean on all touched files.

Could a maintainer approve the workflow run when convenient? (Checks are pending approval for fork PRs.)