refactor: clean up legacy load_state_dict for linear layers#503
Merged
refactor: clean up legacy load_state_dict for linear layers#503
Conversation
There was a problem hiding this comment.
Pull Request Overview
This PR refactors the legacy load_state_dict method for linear layers by removing the transform function variant and updating method calls to use simplified loading mechanisms. The changes also improve code structure by using structured bindings for QKV projections and updating parameter registration to use sharded parameters.
- Removes the legacy
load_state_dictmethod with transform functions from linear layer implementations - Updates QKV projection calls across multiple model architectures to use structured bindings instead of array indexing
- Refactors parameter registration in quantized linear layers to use
register_sharded_parameterinstead ofregister_parameter
Reviewed Changes
Copilot reviewed 11 out of 11 changed files in this pull request and generated no comments.
Show a summary per file
| File | Description |
|---|---|
| src/quantization/qlinear_impl_test.cpp | Updates test calls to use load() and verify() methods instead of legacy state dict methods |
| src/quantization/qlinear_impl.cpp | Replaces parameter registration calls with sharded parameter registration including rank and world_size |
| src/models/meta/llama.h | Replaces QKV array indexing with structured binding for cleaner code |
| src/models/google/gemma2.h | Replaces QKV array indexing with structured binding for cleaner code |
| src/models/google/gemma.h | Replaces QKV array indexing with structured binding for cleaner code |
| src/models/alibaba/qwen2.h | Replaces QKV array indexing with structured binding for cleaner code |
| src/layers/qkv_linear_test.cpp | Updates test to use structured binding for QKV outputs |
| src/layers/qkv_linear.h | Updates forward method signature to return tuple instead of vector |
| src/layers/linear_impl.h | Removes legacy transform function overload declaration |
| src/layers/linear_impl.cpp | Removes legacy transform function implementation and simplifies loading |
| src/layers/linear.h | Removes virtual transform function method from base class |
Tip: Customize your code reviews with copilot-instructions.md. Create the file or learn how to get started.
guocuimi
force-pushed
the
linear_refactor
branch
from
3c883b4 to
31afc66
Compare
guocuimi
commented
Oct 7, 2025
src/models/meta/llama.h
Outdated
| torch::Tensor forward(torch::Tensor x) { | ||
| const auto gate_up = gate_up_proj_(x); | ||
| return down_proj_(act_func_(gate_up[0]) * gate_up[1]); | ||
| // const auto gate_up = gate_up_proj_(x); |
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
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.
No description provided.