Refactor update_block_size_for_backend to reuse vLLM's base implementation

[Alex-ai-future]

PR: Refactor update_block_size_for_backend to reuse vLLM's base implementation

Description

This PR refactors MetalPlatform.update_block_size_for_backend() to delegate to vLLM's base implementation, reducing code duplication and improving maintainability.

Key Changes

1. Override _find_non_ssm_backend() - Returns a synthetic MetalBackend class with get_supported_kernel_block_sizes() returning [MultipleOf(32)], enabling vLLM's base implementation to calculate block_size with Metal-specific kernel alignment.

2. Simplify update_block_size_for_backend() - Now delegates to super().update_block_size_for_backend() for all block_size calculations, keeping only Metal-specific adjustments (ensuring block_size is a multiple of 32 for paged attention).

3. Add manual_seed_all() stub - Fixes NotImplementedError during worker initialization.

4. Update tests - Refactored to test the new implementation, focusing on Metal-specific behavior rather than mocking vLLM internals.

Benefits

• ~340 lines of test code removed - Tests now focus on Metal-specific logic
• ~80 lines of implementation code removed - Reuses vLLM's _align_hybrid_block_size() logic
• Better maintainability - Future changes to vLLM's hybrid model handling automatically apply to vllm-metal
• Clearer separation - vllm-metal only defines Metal-specific kernel_block_alignment_size=32

Testing

• Existing unit tests updated and passing
• Ruff lint and format checks passing

Files Changed

• vllm_metal/platform.py - Core implementation
• tests/test_platform_update_block_size.py - Unit tests

Code Statistics

• Deleted: 623 lines
• Added: 218 lines
• Net reduction: 405 lines

[Alex-ai-future]

It looks like the test got stuck for some reason. @WindChimeRan

[ericcurtin]

This needs work:

super().update_block_size_for_backend() doesn't exist in vLLM Platform, _find_non_ssm_backend is never called by vLLM, imports from non-existent module/class (MultipleOf, wrong AttentionBackend path), massive test coverage regression (521→103 lines), unrelated manual_seed_all addition, Metal-specific adjustment code is unreachable after super() failure.

Btw, qwen coder free tier is very good at generating slop, so be careful, it's still useful, but have to keep a very close eye on it, requires a lot of steering

[Alex-ai-future]

This needs work:

super().update_block_size_for_backend() doesn't exist in vLLM Platform, _find_non_ssm_backend is never called by vLLM, imports from non-existent module/class (MultipleOf, wrong AttentionBackend path), massive test coverage regression (521→103 lines), unrelated manual_seed_all addition, Metal-specific adjustment code is unreachable after super() failure.

Btw, qwen coder free tier is very good at generating slop, so be careful, it's still useful, but have to keep a very close eye on it, requires a lot of steering

This may be because I use the latest version of vllm code which adds some new methods to platform. In the current vllm version （0.19.0）, the update function does not need to be calculated(I find surprisedly), because it has been updated once by another place before (during init configuration), so the current change is actually compatible with both the current version and the latest version. Maybe I need to explain the situation with comments. What do you think?

Or can you keep this pr to till the next vllm upgrade?

In fact, I actually use qwen coder after debugging and troubleshooting the problem. Of course, I will be more careful. Since my understanding still needs to be improved, thank you for your suggestions.

[Alex-ai-future]

I have updated some comments to clarify the version-related issues.

Compatibility note:

• vLLM 0.19.0 (currently supported): Block size is configured during
  VllmConfig initialization, so this method is not strictly required.
  However, we still implement it for forward compatibility.
• Latest vLLM: The block_size calculation logic was moved to use
  Platform.update_block_size_for_backend() in the base class.

This implementation delegates to super().update_block_size_for_backend()
to support both versions. Our overridden _find_non_ssm_backend provides
the Metal-specific kernel block alignment (MultipleOf(32)) that the base
class uses for hybrid model block_size calculation.

[WindChimeRan]

This PR is coupled with vllm version (post-0.19.0). I want to defer the review to next stable version we support (e.g., 0.20.0), so the reviewer don't have to check both versions of vllm.

[Alex-ai-future]

This PR is coupled with vllm version (post-0.19.0). I want to defer the review to next stable version we support (e.g., 0.20.0), so the reviewer don't have to check both versions of vllm.

I consider this to be entirely reasonable; the only drawback is that the current vLLM Metal implementation is currently unusable with the latest version of vLLM due to a lack of necessary implementation methods. Consequently, debugging vLLM might be a bit troublesome.

[ericcurtin]

@Alex-ai-future it's kinda on you if you use an in-development version of vllm , the latest release is

https://github.com/vllm-project/vllm/releases/tag/v0.19.0

any commit after that point shouldn't be under consideration right now

[Alex-ai-future]

@Alex-ai-future it's kinda on you if you use an in-development version of vllm , the latest release is

https://github.com/vllm-project/vllm/releases/tag/v0.19.0

any commit after that point shouldn't be under consideration right now

OK, then my contribution can be regarded as an attempt or reference. If you have any other requirements, I am willing to cooperate.

[ericcurtin]

Closing for now