perf: optimize rejection sampling triton kernel #25791
Open
Conversation
This file contains hidden or 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
happierpig
requested review from
WoosukKwon,
alexm-redhat,
comaniac,
njhill,
robertgshaw2-redhat and
ywang96
as code owners
![gemini-code-assist[bot]](https://avatars.githubusercontent.com/in/956858?s=60&v=4){kind=small}
There was a problem hiding this comment.
Code Review
This pull request optimizes the sample_recovered_tokens_kernel Triton kernel by unrolling the computation over the vocabulary size. This is a good performance optimization for large vocabularies. The implementation of the online maximum calculation is correct. However, I've identified a performance issue where a value is redundantly loaded within a loop. My review includes a suggestion to fix this.
Comment on lines
+602
to
+640
| for off in range(0, vocab_size, BLOCK_SIZE): | ||
| vocab_offset = off + tl.arange(0, BLOCK_SIZE) | ||
| if NO_DRAFT_PROBS: | ||
| draft_token_id = tl.load(draft_token_ids_ptr + start_idx + pos) | ||
| prob = tl.load( | ||
| target_probs_ptr + (start_idx + pos) * vocab_size + | ||
| vocab_offset, | ||
| mask=((vocab_offset < vocab_size) & | ||
| (vocab_offset != draft_token_id)), | ||
| other=0, | ||
| ) | ||
| else: | ||
| draft_prob = tl.load( | ||
| draft_probs_ptr + (start_idx + pos) * vocab_size + | ||
| vocab_offset, | ||
| mask=vocab_offset < vocab_size, | ||
| other=0, | ||
| ) | ||
| target_prob = tl.load( | ||
| target_probs_ptr + (start_idx + pos) * vocab_size + | ||
| vocab_offset, | ||
| mask=vocab_offset < vocab_size, | ||
| other=float("-inf")) | ||
| recovered_id = tl.argmax(prob / q, axis=-1) | ||
| tl.store(output_token_ids_ptr + start_idx + pos, recovered_id) | ||
| other=0, | ||
| ) | ||
| prob = tl.maximum(target_prob - draft_prob, 0) | ||
|
|
||
| q = tl.load( | ||
| q_ptr + req_idx * vocab_size + vocab_offset, | ||
| mask=vocab_offset < vocab_size, | ||
| other=float("-inf"), | ||
| ) | ||
| scores = prob / q | ||
| local_val = tl.max(scores, axis=-1) | ||
| local_idx = tl.argmax(scores, axis=-1) + off | ||
|
|
||
| # update global max | ||
| better = local_val > max_val | ||
| max_val = tl.where(better, local_val, max_val) | ||
| max_idx = tl.where(better, local_idx, max_idx) |
There was a problem hiding this comment.
For performance, draft_token_id should be loaded only once before the loop since its value doesn't change across iterations. Loading it inside the loop results in redundant reads from global memory, which can negatively impact kernel performance.
if NO_DRAFT_PROBS:
draft_token_id = tl.load(draft_token_ids_ptr + start_idx + pos)
for off in range(0, vocab_size, BLOCK_SIZE):
vocab_offset = off + tl.arange(0, BLOCK_SIZE)
if NO_DRAFT_PROBS:
prob = tl.load(
target_probs_ptr + (start_idx + pos) * vocab_size +
vocab_offset,
mask=((vocab_offset < vocab_size) &
(vocab_offset != draft_token_id)),
other=0,
)
else:
draft_prob = tl.load(
draft_probs_ptr + (start_idx + pos) * vocab_size +
vocab_offset,
mask=vocab_offset < vocab_size,
other=0,
)
target_prob = tl.load(
target_probs_ptr + (start_idx + pos) * vocab_size +
vocab_offset,
mask=vocab_offset < vocab_size,
other=0,
)
prob = tl.maximum(target_prob - draft_prob, 0)
q = tl.load(
q_ptr + req_idx * vocab_size + vocab_offset,
mask=vocab_offset < vocab_size,
other=float("-inf"),
)
scores = prob / q
local_val = tl.max(scores, axis=-1)
local_idx = tl.argmax(scores, axis=-1) + off
# update global max
better = local_val > max_val
max_val = tl.where(better, local_val, max_val)
max_idx = tl.where(better, local_idx, max_idx)
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
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.
Purpose
sample_recovered_tokens_kernelimplementation by unrolling CTA over thevocab_sizedimension.num_warps, could be helpful for small batch sizes.Test Plan
Test Result
Essential Elements of an Effective PR Description Checklist
supported_models.mdandexamplesfor a new model.