[Model] Consolidate Deepseek-MoE implementation with DeepSeek-v2

[Isotr0py]

Purpose

• Fix [Feature]: Enable DeepSeek-OCR pluginization #28096
• Since the difference between ds-moe and ds-v2 is only the MHA/MLA, we can consolidate them.

Test Plan

python examples/offline_inference/vision_language.py -m deepseek_ocr

python examples/offline_inference/basic/generate.py --model deepseek-ai/DeepSeek-V2-Lite-Chat

Test Result

Have confirmed CPU/XPU can run DeepSeek-COR now, and Deepseek-V2 isn't affected.

---

Essential Elements of an Effective PR Description Checklist

• The purpose of the PR, such as "Fix some issue (link existing issues this PR will resolve)".
• The test plan, such as providing test command.
• The test results, such as pasting the results comparison before and after, or e2e results
• (Optional) The necessary documentation update, such as updating supported_models.md and examples for a new model.
• (Optional) Release notes update. If your change is user facing, please update the release notes draft in the Google Doc.

[mergify]

This pull request has merge conflicts that must be resolved before it can be
merged. Please rebase the PR, @Isotr0py.

https://docs.github.com/en/pull-requests/collaborating-with-pull-requests/working-with-forks/syncing-a-fork

[chatgpt-codex-connector]

💡 Codex Review

vllm/vllm/model_executor/models/deepseek_v2.py

Lines 1277 to 1283 in 1ca29da

	class DeepseekV2ForCausalLM(nn.Module, SupportsPP, MixtureOfExperts, SupportsLoRA):
	packed_modules_mapping = {
	"gate_up_proj": ["gate_proj", "up_proj"],
	}
	def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
	super().__init__()

Reintroduce qkv packed mapping for Deepseek LoRA modules

The old DeepseekForCausalLM exposed a packed_modules_mapping entry for "qkv_proj": ["q_proj", "k_proj", "v_proj"] so LoRA and quantization code could translate user‑facing module names to the packed QKVParallelLinear weights. After consolidating the Deepseek implementation into deepseek_v2.py, the new DeepseekForCausalLM simply inherits DeepseekV2ForCausalLM, whose mapping only includes gate_up_proj (and optionally fused_qkv_a_proj). As a result, LoRA loaders now have no mapping for q_proj, k_proj, or v_proj when targeting Deepseek MoE models (see vllm/lora/models.py for how these mappings are consumed), so adapters that previously worked will fail to register or load. Consider restoring the qkv_proj entry when self.use_mha is true so LoRA and packed quantization continue to work for the original Deepseek models.

ℹ️ About Codex in GitHub

Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".

[jikunshang]

LGTM. thanks for refactoring!

[mergify]

This pull request has merge conflicts that must be resolved before it can be
merged. Please rebase the PR, @Isotr0py.

https://docs.github.com/en/pull-requests/collaborating-with-pull-requests/working-with-forks/syncing-a-fork

[mergify]

This pull request has merge conflicts that must be resolved before it can be
merged. Please rebase the PR, @Isotr0py.

https://docs.github.com/en/pull-requests/collaborating-with-pull-requests/working-with-forks/syncing-a-fork

[gcanlin]

LGTM. Thanks for the good work! I have validated Deepseek-OCR on Ascend NPU (910B) using the latest vllm-ascend main branch plus this PR, and the results are successful.

[sixgod-666]

LGTM. Thanks for the good work! I have validated Deepseek-OCR on Ascend NPU (910B) using the latest vllm-ascend main branch plus this PR, and the results are successful.

How can I obtain a mirror environment that can successfully run DeepSeek-OCR? I tried the vllm-ascend-main image and found that its vllm version is v0.11.0.