[Bugfix] Fix PP for ChatGLM and Molmo, and weight loading for Qwen2.5-Math-RM

[DarkLight1337]

#9242 accidentally removed PP support from ChatGLM. On the other hand, MolmoModel already supports PP but the top-level MolmoForCausalLM does not. This PR fixes both of these issues.

The input processors of these models and Qwen2-VL have also been cleaned up a bit.

Also fixes #9090 (comment)

[github-actions]

👋 Hi! Thank you for contributing to the vLLM project.
Just a reminder: PRs would not trigger full CI run by default. Instead, it would only run fastcheck CI which starts running only a small and essential subset of CI tests to quickly catch errors. You can run other CI tests on top of those by going to your fastcheck build on Buildkite UI (linked in the PR checks section) and unblock them. If you do not have permission to unblock, ping simon-mo or khluu to add you in our Buildkite org.

Once the PR is approved and ready to go, your PR reviewer(s) can run CI to test the changes comprehensively before merging.

To run CI, PR reviewers can do one of these:

• Add ready label to the PR
• Enable auto-merge.

🚀

[DarkLight1337]

Sorry for breaking your code! @lixiaolx can you test out this PR and see if you can load Qwen/Qwen2.5-Math-RM-72B now?

[Isotr0py]

GLM4-V and Molmo PP tests all passed on my devices. LGTM once Qwen2-RM is also confirmed to work.

[lixiaolx]

Sorry for breaking your code! @lixiaolx can you test out this PR and see if you can load Qwen/Qwen2.5-Math-RM-72B now?

@DarkLight1337 , I tried adding lm_head filtering to the model.py file of qwen_rm.
loader = AutoWeightsLoader(self,skip_prefixes=["lm_head"])
Now I can load the model, but the final results of my test reference are all nan #8896. I tried printing the weights of the model, but it seemed that the weights were not loaded correctly.

add print code in
https://github.com/vllm-project/vllm/blob/main/vllm/model_executor/model_loader/loader.py#L413
for name, param in model.named_parameters(): print(f"Parameter '{name}' =====tensor=======: {param}")

[DarkLight1337]

Sorry for breaking your code! @lixiaolx can you test out this PR and see if you can load Qwen/Qwen2.5-Math-RM-72B now?

@DarkLight1337 , I tried adding lm_head filtering to the model.py file of qwen_rm.
loader = AutoWeightsLoader(self,skip_prefixes=["lm_head"])
Now I can load the model, but the final results of my test reference are all nan #8896. I tried printing the weights of the model, but it seemed that the weights were not loaded correctly.

add print code in
https://github.com/vllm-project/vllm/blob/main/vllm/model_executor/model_loader/loader.py#L413
for name, param in model.named_parameters(): print(f"Parameter '{name}' =====tensor=======: {param}")

Can you add print statements inside AutoWeightLoader and list out the weight names in the iterator? This should provide enough information for me to figure out the discrepancies and fix them.

[DarkLight1337]

Sorry for breaking your code! @lixiaolx can you test out this PR and see if you can load Qwen/Qwen2.5-Math-RM-72B now?

@DarkLight1337 , I tried adding lm_head filtering to the model.py file of qwen_rm.
loader = AutoWeightsLoader(self,skip_prefixes=["lm_head"])
Now I can load the model, but the final results of my test reference are all nan #8896. I tried printing the weights of the model, but it seemed that the weights were not loaded correctly.
add print code in
https://github.com/vllm-project/vllm/blob/main/vllm/model_executor/model_loader/loader.py#L413
for name, param in model.named_parameters(): print(f"Parameter '{name}' =====tensor=======: {param}")

Can you add print statements inside AutoWeightLoader and list out the weight names in the iterator? This should provide enough information for me to figure out the discrepancies and fix them.

Any update? Otherwise, I'll merge this PR first and fix the model in another PR.

[DarkLight1337]

Sorry for breaking your code! @lixiaolx can you test out this PR and see if you can load Qwen/Qwen2.5-Math-RM-72B now?

@DarkLight1337 , I tried adding lm_head filtering to the model.py file of qwen_rm.
loader = AutoWeightsLoader(self,skip_prefixes=["lm_head"])
Now I can load the model, but the final results of my test reference are all nan #8896. I tried printing the weights of the model, but it seemed that the weights were not loaded correctly.
add print code in
https://github.com/vllm-project/vllm/blob/main/vllm/model_executor/model_loader/loader.py#L413
for name, param in model.named_parameters(): print(f"Parameter '{name}' =====tensor=======: {param}")

Can you add print statements inside AutoWeightLoader and list out the weight names in the iterator? This should provide enough information for me to figure out the discrepancies and fix them.

Any update? Otherwise, I'll merge this PR first and fix the model in another PR.

I think I fixed the problem, so I'm going to merge this now. Please tell me and share the logs (set the environment variable VLLM_LOGGING_LEVEL=DEBUG) if it still doesn't work.