[Bugfix] Fix marlin kernel crash on H100 #4218
Merged
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
mgoin
reviewed
Apr 20, 2024
|
I tried various modifications to the PTX to keep the cache-hint, but it did not work. |
pcmoritz
approved these changes
Apr 24, 2024
There was a problem hiding this comment.
Thanks for fixing this, I validated the fix with the reproduction in neuralmagic#187. Always great to see fixes that make things simpler ❤️
xjpang
pushed a commit
to xjpang/vllm
that referenced
this pull request
Apr 25, 2024
This PR addresses the Marlin kernel H100 crash that was reported here: neuralmagic#187. The reason for the crash was the inline PTX assembly that introduced the async_copy with streaming behavior. The solution is to use the more standard PTX for async_copy (without the fractional L2 policy for "evict_first"). There is no performance difference between standard async_copy PTX and the previous one.
robertgshaw2-neuralmagic
pushed a commit
to neuralmagic/nm-vllm
that referenced
this pull request
Apr 26, 2024
This PR addresses the Marlin kernel H100 crash that was reported here: neuralmagic#187. The reason for the crash was the inline PTX assembly that introduced the async_copy with streaming behavior. The solution is to use the more standard PTX for async_copy (without the fractional L2 policy for "evict_first"). There is no performance difference between standard async_copy PTX and the previous one.
3 tasks
alexeykondrat
pushed a commit
to alexeykondrat/ci-vllm
that referenced
this pull request
May 1, 2024
This PR addresses the Marlin kernel H100 crash that was reported here: neuralmagic#187. The reason for the crash was the inline PTX assembly that introduced the async_copy with streaming behavior. The solution is to use the more standard PTX for async_copy (without the fractional L2 policy for "evict_first"). There is no performance difference between standard async_copy PTX and the previous one.
z103cb
pushed a commit
to z103cb/opendatahub_vllm
that referenced
this pull request
May 7, 2024
This PR addresses the Marlin kernel H100 crash that was reported here: neuralmagic#187. The reason for the crash was the inline PTX assembly that introduced the async_copy with streaming behavior. The solution is to use the more standard PTX for async_copy (without the fractional L2 policy for "evict_first"). There is no performance difference between standard async_copy PTX and the previous one.
mgoin
added a commit
to neuralmagic/nm-vllm
that referenced
this pull request
May 16, 2024
The reason for the crash was the inline PTX assembly that introduced the async_copy with streaming behavior. The solution is to use the more standard PTX for async_copy (without the fractional L2 policy for "evict_first"). There is no performance difference between standard async_copy PTX and the previous one. Ported from dense marlin: vllm-project#4218
3 tasks
|
Is there any data to prove that there is no difference for l2 cache hint method in performance? But why is this method introduced in the marlin original paper? |
Temirulan
pushed a commit
to Temirulan/vllm-whisper
that referenced
this pull request
Sep 6, 2024
This PR addresses the Marlin kernel H100 crash that was reported here: neuralmagic#187. The reason for the crash was the inline PTX assembly that introduced the async_copy with streaming behavior. The solution is to use the more standard PTX for async_copy (without the fractional L2 policy for "evict_first"). There is no performance difference between standard async_copy PTX and the previous one.
Hello, Is there any data to prove that there is no difference for l2 cache hint method in performance? I use a100 to run bench.py on marlin repo and it showes that on models other than llama 7B, l2 cache hint has performance improvement. Although bench.py actually only does matmul operation for simulation but does not have a full inference procedure. |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
This PR addresses the Marlin kernel H100 crash that was reported here: neuralmagic#187.
The reason for the crash was the inline PTX assembly that introduced the async_copy with streaming behavior. The solution is to use the more standard PTX for async_copy (without the fractional L2 policy for "evict_first"). There is no performance difference between standard async_copy PTX and the previous one.