Skip to content

TF upgrade to 5.19.0 - #1287

Open
Mamta Singh (quic-mamta) wants to merge 25 commits into
qualcomm:mainfrom
quic-mamta:tf_upgrade
Open

TF upgrade to 5.19.0#1287
Mamta Singh (quic-mamta) wants to merge 25 commits into
qualcomm:mainfrom
quic-mamta:tf_upgrade

Conversation

@quic-mamta

@quic-mamta Mamta Singh (quic-mamta) commented Aug 26, 2026 •
edited
Loading

Copy link
Copy Markdown
Contributor

Summary

This PR upgrades QEfficient’s Transformers dependency from 5.5.4 to 5.19.0 and updates the affected model wrappers/tests for API changes introduced in the newer Transformers release.

Changes

  • Updated the Transformers pin to 5.19.0.
  • Added compatibility for renamed/moved Transformers classes:
    • CLIPTextTransformer fallback to CLIPTextModel
    • GraniteMoE expert/router rename handling
  • Fixed Gemma4 heterogeneous config handling:
    • Avoid direct global access to per-layer attributes like head_dim/global_head_dim
    • Resolve per-layer config values when building dummy KV cache tensors
    • Updated Gemma4 accuracy tests to avoid unsafe hasattr(...) on heterogeneous configs
  • Updated Qwen3.5 and Qwen3.5-MoE wrappers for Transformers 5.19.0 recurrent attention mask/layer-type behavior.
  • Fixed GLM4-MoE router reshaping to use hidden_dim.
  • Updated transform coverage for renamed optimized MoE components.

@quic-mamta
Mamta Singh (quic-mamta) marked this pull request as draft August 26, 2026 08:34
@quic-mamta
Mamta Singh (quic-mamta) force-pushed the tf_upgrade branch 3 times, most recently from 6f095ae to ab11e35 Compare August 26, 2026 10:41
ochougul
Onkar Chougule (ochougul) reviewed Aug 26, 2026
View reviewed changes
Comment thread QEfficient/transformers/models/clip/modeling_clip.py Outdated
@quic-mamta
Mamta Singh (quic-mamta) marked this pull request as ready for review August 31, 2026 06:39
@quic-mamta

Mamta Singh (quic-mamta) commented Sep 22, 2026

Copy link
Copy Markdown
Contributor Author

CI-Ready

1 similar comment
@quic-mamta

Mamta Singh (quic-mamta) commented Sep 27, 2026

Copy link
Copy Markdown
Contributor Author

CI-Ready

@quic-mamta

Mamta Singh (quic-mamta) commented Sep 28, 2026

Copy link
Copy Markdown
Contributor Author

CI-Ready

@quic-mamta
Mamta Singh (quic-mamta) force-pushed the tf_upgrade branch 2 times, most recently from 9a12fad to 3b44e97 Compare September 29, 2026 08:36
@quic-mohmeh quic-mohmeh mentioned this pull request Oct 5, 2026
Merged
@quic-mamta

Mamta Singh (quic-mamta) commented Oct 14, 2026

Copy link
Copy Markdown
Contributor Author

CI-Ready

@quic-mamta Mamta Singh (quic-mamta) changed the title TF upgrade to 5.15.1 TF upgrade to 5.19.0 Oct 16, 2026
@quic-mamta
Mamta Singh (quic-mamta) force-pushed the tf_upgrade branch 4 times, most recently from 35ee37c to ddd401c Compare October 23, 2026 17:34
@quic-mamta
upgrade tf to 5.15.1
756b1dc 
Signed-off-by: Mamta Singh <mamtsing@qti.qualcomm.com>
@quic-mamta
update glm, qwen3_5 and granitemoe
49301d9 
Signed-off-by: Mamta Singh <mamtsing@qti.qualcomm.com>
@quic-mamta
fix diffusers and gemma4
833d4b6 
Signed-off-by: Mamta Singh <mamtsing@qti.qualcomm.com>
Mamta Singh (quic-mamta) and others added 18 commits October 23, 2026 23:22
@quic-mamta
fix subfunctions
5a4cf28 
Signed-off-by: Mamta Singh <mamtsing@qti.qualcomm.com>
@quic-mamta
fix unit tests
7443f44 
Signed-off-by: Mamta Singh <mamtsing@qti.qualcomm.com>
@quic-mamta
ruff check
acb32c9 
Signed-off-by: Mamta Singh <mamtsing@qti.qualcomm.com>
@quic-mamta
lint/format
fd9e382 
Signed-off-by: Mamta Singh <mamtsing@qti.qualcomm.com>
@quic-mamta
fix test failures
1e7f455 
Signed-off-by: Mamta Singh <mamtsing@qti.qualcomm.com>
@quic-mamta
fix gemma4
c6b7520 
Signed-off-by: Mamta Singh <mamtsing@qti.qualcomm.com>
@quic-mamta
fix gemma3 llava golden output
457e4fe 
Signed-off-by: Mamta Singh <mamtsing@qti.qualcomm.com>
@quic-vishali @quic-mamta
Logger Module For Efficient Transformers (qualcomm#1282)
59183ab 
Adds a structured logging module (`QEFFLogger`) to the QEfficient
library,
based on
[qualcomm#696](qualcomm#696).

- `QEFFLogger` is a singleton logger with JSON formatter and rotating
file handler.
- Log level controlled project-wide via `QEFF_LOG_LEVEL` env var
(default: INFO); log file path
via `QEFF_LOG_PATH` or a timestamped default under
`~/.cache/qefficient_logs/`.
- Namespace support per module (`INFRA`, `MODEL`, `FT`) for targeted
filtering and post-processing.
- All existing `logging.getLogger(__name__)` and module-level `logger`
imports migrated to
    `QEFFLogger.get_logger(namespace)`.
- Jenkinsfile: all CI stages set
`QEFF_LOG_PATH=/efficient-transformers/Qeff_logs`.
  - New unit tests in `tests/utils/test_logger.py`.
- Generates process-aware filenames to avoid collisions between parallel
workers:
    `QEFF_YYYYMMDD_HHMMSS_PID.log`.

---------

Signed-off-by: Vishali Senthilkumar <vishsent@qti.qualcomm.com>
@quic-mamta
fix unit tests
ecb97f1 
Signed-off-by: Mamta Singh <mamtsing@qti.qualcomm.com>
@quic-mamta
update huggingface-hub version for vllm
6327b35 
Signed-off-by: Mamta Singh <mamtsing@qti.qualcomm.com>
@quic-mamta
fix mpnet
5324fc2 
Signed-off-by: Mamta Singh <mamtsing@qti.qualcomm.com>
@quic-mamta
Upgrade TF to 5.19.0
35f50b5 
Signed-off-by: Mamta Singh <mamtsing@qti.qualcomm.com>
@quic-mamta
fix gemma4
c1280b6 
Signed-off-by: Mamta Singh <mamtsing@qti.qualcomm.com>
@quic-mamta
fix qwen 3_5 moe
982bed9 
Signed-off-by: Mamta Singh <mamtsing@qti.qualcomm.com>
@quic-mamta
fix rotary embedding and mdp strategy
c6d93b8 
Signed-off-by: Mamta Singh <mamtsing@qti.qualcomm.com>
@quic-amitraj @quic-akuruvil
@quic-dhirajku @tchawada @quic-mamta
Enable weight-free export for supported VLM families (qualcomm#1440)
439db84 
Adds weight-free export, compile, and runtime support for
image-text-to-text
models. The implementation uses Dynamo-based ONNX export and
externalized
  checkpoint weights, with dual-QPC VLM execution.

  ### Supported VLM families

  - Qwen2.5-VL
  - Qwen3-VL
  - Qwen3-VL-MoE
  - Qwen3.5
  - Qwen3.5-MoE
  - Gemma4, including dense and MoE configurations

  ## What changed

  - Add weight-free VLM construction and dual-QPC routing in
    `QEFFAutoModelForImageTextToText`.
- Add checkpoint-key resolution and weight transforms for VLM and
Gemma4/MoE
    checkpoints.
- Extend Dynamo/ONNX export handling for repeated decoder subgraphs and
    external weight specifications.
  - Add combined and disaggregated vision, prefill, and decode flows.
- Add examples and CPU/QAIC coverage for weight-free VLM export,
compile, and
    HF-to-QAIC token parity.

  ## Constraints and known limitations

- Weight-free VLM execution requires dual-QPC mode (`kv_offload=True`).
  - Single-QPC weight-free export is unsupported.
  - `weight_free=True` and `layerwise=True` are mutually exclusive.
- Qwen2.5-VL multi-specialization vision compilation remains deferred
due to a
    QAIC Conv3D bias-shape error.
  - Gemma4 multi-specialization is not supported yet.
- Gemma4 continuous batching can fail to compile with `full_batch_size >
2`;
    CB + disaggregated compilation is also currently limited.

  ## Validation

  - Not run during this PR preparation.
- Added/updated coverage includes CPU export structure, dual-QPC
compilation,
    CCL, disaggregated flows, and QAIC HF-token parity.

---------
Signed-off-by: Ann Kuruvilla <akuruvil@qti.qualcomm.com>
Signed-off-by: Amit Raj <amitraj@qti.qualcomm.com>
Signed-off-by: Dhiraj Kumar Sah <dhirajku@qti.qualcomm.com>
Signed-off-by: Tanisha Chawada <tchawada@qti.qualcomm.com>
Co-authored-by: Ann Kuruvilla <akuruvil@qti.qualcomm.com>
Co-authored-by: Dhiraj Kumar Sah <dhirajku@qti.qualcomm.com>
Co-authored-by: Tanisha Chawada <tchawada@qti.qualcomm.com>
@quic-mamta
fix weight free tests
2045ee3 
Signed-off-by: Mamta Singh <mamtsing@qti.qualcomm.com>
@quic-mamta
fix gemma3 and qwen_2_5_vl
82ccd4b 
Signed-off-by: Mamta Singh <mamtsing@qti.qualcomm.com>
@quic-mamta
lint/format
12e0a38 
Signed-off-by: Mamta Singh <mamtsing@qti.qualcomm.com>
@quic-mamta
fix gemma3
b768824 
Signed-off-by: Mamta Singh <mamtsing@qti.qualcomm.com>
@quic-mamta
fix gemma4
c25c7b1 
Signed-off-by: Mamta Singh <mamtsing@qti.qualcomm.com>

This branch has not been deployed

No deployments
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

@quic-rishinr Rishin Raj (quic-rishinr) Awaiting requested review from quic-rishinr
@ochougul Onkar Chougule (ochougul) Awaiting requested review from ochougul

Assignees

@quic-mamta Mamta Singh (quic-mamta)

Labels

ci_ready

Projects

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.

4 participants

@quic-mamta @ochougul @quic-vishali @quic-amitraj