Chunk weight quantization to bound fp32 VRAM spikes - #1659
Open
dxqb wants to merge 1 commit into
Open
Conversation
Quantizing a W8A8 weight upcast the whole tensor to fp32, and the non-in-place `mul` made a second full-size copy, so one large layer transiently needed several times its own size in VRAM. On small GPUs that was enough to OOM during quantization. `quantize_int8_tensorwise_chunked` / `quantize_fp8_tensorwise_chunked` quantize the weight in row-blocks of `_QUANTIZE_CHUNK_ELEMENTS`, bounding the fp32 transient to one block; `LinearW8A8.quantize` uses them. The tensorwise scale now comes from `torch.aminmax` instead of `abs().max()`, dropping another full-tensor copy. The shared `quantize_int8` / `quantize_fp8` cores take an explicit fp32 copy and then scale in place, so an already-fp32 caller tensor is never clobbered. Output is unchanged, bit for bit. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
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
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Summary
Quantizing a W8A8 weight upcast the whole tensor to fp32, and the non-in-place
mulmade a second full-size copy, so one large layer transiently needed several times its own size in VRAM. On small GPUs that was enough to OOM during quantization.Test plan
pre-commit run --all-filespassesAI assistance