PR #39480: [CI/Build][Model][DeepSeek] Add e2e correctness tests for compile fusions

• Repository: vllm-project/vllm
• Author: xikronz
• State: open | merged: False
• Link: https://github.com/vllm-project/vllm/pull/39480

Description (problem / solution / changelog)

Purpose

WIP for issue #39428 part 1: loading partial layers in deepseek

• fixes --hf-overrides.num_hidden_layers to bypass KeyErrors in DeepSeek_v2 as a proof of concept during weight loading when layers are removed via hf_overrides. This is a prerequisite for running few-layer correctness tests on all models in tests/compile/fusions_e2e/models.py.
• do the same for all three deepseek models in tests/compile/fusions_e2e/models.py.

Test plan

I wrote a quick smoketest script for loding 2 layers of deepseek-coder-v2 to see if the keyerror gets triggered for loading 2 layers of deepseek-ai/DeepSeek-Coder-V2-Lite-Base and RedHatAI/DeepSeek-Coder-V2-Lite-Instruct-FP8 which uses deepseek_v2.py and it works!

whereas if I switch back to main branch and run the same smoketest on the same models, we get the expected key error: <img width="1786" height="313" alt="Screenshot from 2026-04-16 00-23-20" src="https://github.com/user-attachments/assets/15f3d088-28b0-4616-bba9-bef2f1ebd761" />

• 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.

llm = LLM(
  model="deepseek-ai/DeepSeek-Coder-V2-Lite-Base",
  tensor_parallel_size=1,
  max_model_len=128,
  enforce_eager=True,
  hf_overrides={"num_hidden_layers": 2},
  trust_remote_code=True,
)
outputs = llm.generate(["Hello, my name is"], SamplingParams(max_tokens=10))
for output in outputs:
  print(f"Prompt: {output.prompt}")
  print(f"Output: {output.outputs[0].text}")

Changed files

• vllm/model_executor/models/deepseek_v2.py (modified, +13/-5)