Skip to content

Experimental/wf kv blocking - #3

Open
smedhe wants to merge 25 commits into
mainfrom
experimental/wf_kv_blocking
Open

smedhe wants to merge 25 commits into
mainfrom
experimental/wf_kv_blocking

Conversation

@smedhe

@smedhe smedhe commented Sep 15, 2026

Copy link
Copy Markdown
Owner

No description provided.

ochougul and others added 25 commits September 1, 2026 10:42
Signed-off-by: Amar <amarshar@qti.qualcomm.com>
Signed-off-by: Amit Raj <amitraj@qti.qualcomm.com>
Co-authored-by: Amit Raj <amitraj@qti.qualcomm.com>
Signed-off-by: Amit Raj <amitraj@qti.qualcomm.com>
- Use `QEfficient.utils._utils.hf_download` instead of calling
`huggingface_hub.snapshot_download` directly in weight-free
  checkpoint resolution.
- Swap the `gpt_oss` weight-free test fixture from the full
`openai/gpt-oss-20b` checkpoint to a lightweight
  `tiny-random/gpt-oss-bf16` stub.
- Remove personal-environment `.gitignore` entries (`env1/`,
`test_models/`, `dbg.log`) — local dev conventions, not
  project-wide.

---------

Signed-off-by: amarshar <amarshar@qti.qualcomm.com>
Signed-off-by: Amit Raj <amitraj@qti.qualcomm.com>
Signed-off-by: Amit Raj <amitraj@qti.qualcomm.com>
Signed-off-by: Amit Raj <amitraj@qti.qualcomm.com>
Signed-off-by: Amit Raj <amitraj@qti.qualcomm.com>
Signed-off-by: Amit Raj <amitraj@qti.qualcomm.com>
Signed-off-by: Amit Raj <amitraj@qti.qualcomm.com>
Signed-off-by: Amit Raj <amitraj@qti.qualcomm.com>
Signed-off-by: Amit Raj <amitraj@qti.qualcomm.com>
Summary

This PR extends the weight-free export pipeline for CausalLM models with
CCL support, fixes several export-path bugs, and
  rounds out the weight-free test suite.

  Changes

  Export pipeline

- utils/cache.py: Add QEFF_CHECKPOINT_HOME env var — controls where
prepared checkpoints (dtype-converted, MoE-restacked)
- utils/torch_patches.py: Add dynamo_invoke_subgraph_fallback_env()
context manager that sets
TORCH_INVOKE_ALLOW_CREATE_FALLBACK=1 for dynamo export; simplify
setattr(module, "forward", ...) to direct assignment
- utils/export_utils.py: Fix dynamo flag inference in export_wrapper
(removed incorrect use_weight_free_export coercion);
  simplify nested with context managers
- base/modeling_qeff.py, exporter/onnx_exporter.py,
exporter/weight_free/export.py,
exporter/weight_free/checkpoint_key_resolver.py,
transformers/models/modeling_auto.py: Weight-free export pipeline
updates
  including tied-weight alias handling and CCL integration

  Test suite
- tests/weight_free/test_ccl.py (new): Weight-free + CCL
compile/generate tests — exports once, compiles twice (auto-CCL and
   explicit comp_ctx_lengths_prefill/decode)
- tests/weight_free/conftest.py: Mark gpt_oss as xfail across all
weight-free tests (SerdeError with dynamo=True +
use_onnx_subfunctions=True); restructure pytest_collection_modifyitems
for single-pass marker application
- tests/weight_free/_helpers.py, test_export.py,
test_continuous_batching.py, test_on_qaic.py: Extended coverage and
helper
  updates

---------

Signed-off-by: Amar <amarshar@qti.qualcomm.com>
Signed-off-by: Amit Raj <amitraj@qti.qualcomm.com>
Co-authored-by: Amit Raj <amitraj@qti.qualcomm.com>
Signed-off-by: Amit Raj <amitraj@qti.qualcomm.com>
Signed-off-by: Amit Raj <amitraj@qti.qualcomm.com>
Signed-off-by: Amit Raj <amitraj@qti.qualcomm.com>
Signed-off-by: Amit Raj <amitraj@qti.qualcomm.com>
Signed-off-by: smedhe <smedhe@qti.qualcomm.com>
…me as dynamo

Signed-off-by: smedhe <smedhe@qti.qualcomm.com>
Signed-off-by: smedhe <smedhe@qti.qualcomm.com>
Signed-off-by: smedhe <smedhe@qti.qualcomm.com>
Signed-off-by: smedhe <smedhe@qti.qualcomm.com>
Signed-off-by: smedhe <smedhe@qti.qualcomm.com>
Signed-off-by: smedhe <smedhe@qti.qualcomm.com>
Signed-off-by: smedhe <smedhe@qti.qualcomm.com>
Signed-off-by: smedhe <smedhe@qti.qualcomm.com>
Signed-off-by: smedhe <smedhe@qti.qualcomm.com>
Signed-off-by: smedhe <smedhe@qti.qualcomm.com>
Signed-off-by: smedhe <smedhe@qti.qualcomm.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

4 participants