[Feature] Add support for Qwen3 Reranker with Sequence Classifier head#730
Open
sigridjineth wants to merge 5 commits intohuggingface:mainfrom
Open
[Feature] Add support for Qwen3 Reranker with Sequence Classifier head#730sigridjineth wants to merge 5 commits intohuggingface:mainfrom
sigridjineth wants to merge 5 commits intohuggingface:mainfrom
Conversation
- Added Qwen3ClassificationHead with flexible tensor loading that handles: - score.weight at top level (for converted Qwen3 rerankers) - classifier.weight/bias patterns for standard models - Updated Qwen3Model and FlashQwen3Model to support classification - Added predict() method implementations for both model variants - Extended Qwen3Config with id2label field for classification - Added test case for Qwen3 reranker models with snapshot The implementation supports Qwen3 models converted to sequence classifiers for reranking tasks (e.g., tomaarsen/Qwen3-Reranker-0.6B-seq-cls). The classification head gracefully handles different tensor naming conventions from various conversion approaches. Tested with both embedding and reranking Qwen3 models. 🤖 Generated with [Claude Code](https://claude.ai/code) Co-Authored-By: Claude <noreply@anthropic.com>
Signed-off-by: Daniel Chalef <131175+danielchalef@users.noreply.github.com>
- Added optional fields `instruction` and `use_template` to `RerankRequest` for custom instructions and template usage. - Updated rerank logic to apply templates conditionally based on model type and user input. - Introduced template formatting for Qwen3 models to improve reranking results. This update allows for more flexible and context-aware reranking capabilities, particularly for Qwen3 models.
…classification head in the Qwen3 model.
Member
|
Hey @sigridjineth, thanks for re-opening this PR! Given the amount of changes this might fall into 1.9.0 rather than 1.8.3, but please let me organize the roadmap and come up with a plan 🤗 |
|
Expected this PR to be merged ：） |
Contributor
|
@alvarobartt Just in case you missed it, any reason this PR hasn't been tagged with the v1.9.0 milestone? |
thomas-hiddenpeak
added a commit
to thomas-hiddenpeak/RMinte-Orin-TEI
that referenced
this pull request
Jan 5, 2026
This commit adapts text-embeddings-inference for NVIDIA Jetson Orin (SM87) and L4 GPU (SM89), and integrates valuable community PRs. Changes: 1. SM87/SM89 CUDA Support - Added compute capability 8.7 and 8.9 support - Modified Dockerfile-cuda-all for multi-arch builds - Updated compute_cap.rs for SM87/89 detection Files: Dockerfile-cuda-all, cuda-all-entrypoint.sh, compute_cap.rs 2. PR huggingface#730: Qwen3 Reranker Support - Added classification head for Qwen3 reranking - Implemented template formatting system for chat-based reranking Files: models/qwen3.rs, core/templates.rs, core/lib.rs 3. PR huggingface#787: Batch Notification Performance Optimization - Implemented AtomicUsize counter for batch processing - Reduced unnecessary notify_one() calls - Only last request in batch triggers thread notification Files: core/infer.rs, router/http/server.rs, router/grpc/server.rs 4. PR huggingface#753: GeLU Activation Consistency Fix - Changed Gelu from approximate (gelu) to exact (gelu_erf) - Added NewGelu variant for backward compatibility Files: layers/linear.rs 5. PR huggingface#790: StaticEmbedding Model Support - Added support for 0_StaticEmbedding/ directory structure - Implemented fallback loading for model weights and tokenizer - Default to Mean pooling for StaticEmbedding models Files: models/static_embedding.rs (new), lib.rs, download.rs, router/lib.rs 6. PR huggingface#746: DebertaV2 Sequence Classification Support - Complete DebertaV2 model implementation - Support for sequence classification tasks (e.g., Llama Prompt Guard) - CPU and CUDA device support Files: models/debertav2.rs (new), lib.rs, models/mod.rs All changes have been tested and compile successfully with: cargo check --all-targets Compilation verified with CUDA support: cargo install --path router -F candle-cuda Target Hardware: NVIDIA Jetson Orin AGX (SM87), L4 GPU (SM89) Date: January 5, 2026
4 tasks
michaelfeil
reviewed
Jan 13, 2026
|
|
||
| // Apply template if needed for Qwen3 rerankers | ||
| let model_id = info.model_id.clone(); | ||
| let use_template = req.use_template.unwrap_or_else(|| { |
Contributor
There was a problem hiding this comment.
I am not sure if this works, since qwen defines the rerank propablity as logprobs of yes / yes + no, which means the classifier of the TEI backend gives you two outputs, while RERANK knows how to work with 1 output only.
michaelfeil
reviewed
Jan 13, 2026
| @@ -0,0 +1,110 @@ | |||
| use std::fmt::Write; | |||
Contributor
There was a problem hiding this comment.
Great way of templating.
michaelfeil
reviewed
Jan 13, 2026
| @@ -1,6 +1,7 @@ | |||
| use crate::flash_attn::flash_attn_varlen; | |||
Contributor
There was a problem hiding this comment.
qwen3 parts LGTM to be, the templating might better be opened as a separate PR?
Author
|
@alvarobartt should I bump this PR to be matched with latest commit? |
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
6 participants
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.
What does this PR do?
Before submitting
instasnapshots?Who can review?
Anyone in the community is free to review the PR once the tests have passed. Feel free to tag
members/contributors who may be interested in your PR.