[Fix] Skip lm_head quantization for q4f16_ft #1731
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
MasterJH5574
approved these changes
Feb 9, 2024
MasterJH5574
pushed a commit
that referenced
this pull request
Feb 17, 2024
This PR reverts the change introduced in #1731 where we skip quantizing `lm_head` to avoid the dynamic vocab issue. This led to performance degradation as pointed out in #1723. Instead, we fall back to `GroupQuantizeLinear` for `lm_head`, which preserves performance and avoids the dynamic vocab size issue. Performance change on RTX 4090 Prefill: throughput: **950.792 --> 973.859 tok/s** total tokens: 7 tok Decode: throughput: **214.372 --> 223.491 tok/s** total tokens: 256 tok
smickey040404
added a commit
to smickey040404/mlc-llm
that referenced
this pull request
Feb 11, 2025
This PR reverts the change introduced in mlc-ai/mlc-llm#1731 where we skip quantizing `lm_head` to avoid the dynamic vocab issue. This led to performance degradation as pointed out in mlc-ai/mlc-llm#1723. Instead, we fall back to `GroupQuantizeLinear` for `lm_head`, which preserves performance and avoids the dynamic vocab size issue. Performance change on RTX 4090 Prefill: throughput: **950.792 --> 973.859 tok/s** total tokens: 7 tok Decode: throughput: **214.372 --> 223.491 tok/s** total tokens: 256 tok
tristankincaid
added a commit
to tristankincaid/mlc-llm
that referenced
this pull request
Feb 16, 2025
This PR reverts the change introduced in mlc-ai/mlc-llm#1731 where we skip quantizing `lm_head` to avoid the dynamic vocab issue. This led to performance degradation as pointed out in mlc-ai/mlc-llm#1723. Instead, we fall back to `GroupQuantizeLinear` for `lm_head`, which preserves performance and avoids the dynamic vocab size issue. Performance change on RTX 4090 Prefill: throughput: **950.792 --> 973.859 tok/s** total tokens: 7 tok Decode: throughput: **214.372 --> 223.491 tok/s** total tokens: 256 tok
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 fixes the issue brought up in #1723.
The issue is due to how the
lm_headof Llama hasout_featuresdefined as atir.Var("vocab_size"). However, in the FT quantization scheme, the linear layer would be quatnzied to shape(in_features, tir.ceildiv(out_features, config.num_elem_per_storage)),. This causes the shape of the parameter to be a FloorDiv operation rather than just atir.Var. Later it would lead to issue during shape-lowering:Instead, we skip quantizing
lm_headhere.