Update: SparseGPT recipes

[rahul-tuli]

Issue

The following test was failing:

FAILED tests/e2e/vLLM/test_vllm.py::TestvLLM_0_tests_e2e_vLLM_configs_sparse2of4_fp8_dynamic_yaml::test_vllm  
ValueError: There is no module or parameter named 'lm_head.bitmask' in LlamaForCausalLM

This issue arose due to recent improvements in SparseGPTModifier, which changed its default behavior. Previously, lm_head was silently ignored, but the new updates no longer do so automatically.

Fix

The fix involves explicitly updating the affected recipes to include the parameter:

ignore: ["re:.*lm_head"]

when all layers are targeted. This ensures that lm_head is properly excluded and prevents the failure.

Example Change

Previously, we relied on regex patterns to target linear layers while ignoring lm_head. The updated configuration now explicitly targets linear layers and ignores lm_head:

-    "targets": ["re:model.layers.\\d+$"],
+    "targets": ["Linear"],
+    "ignore": ["re:.*lm_head"]

This provides a more structured approach and avoids unnecessary regex-based filtering.

Additional Fixes & Improvements

• Removed the deprecated argument sequential_update.
• Updated recipes to use "targets": ["Linear"] instead of regex matching for better clarity and maintainability.
• Raise Warning when lm_head is targetted. Contributed by @kylesayrs

---

• To see the specific tasks where the Asana app for GitHub is being used, see below:
  • https://app.asana.com/0/0/1209368043192615

[github-actions]

👋 Hi! Thank you for contributing to llm-compressor. Please add the ready label when the PR is ready for review.

Note: This is required to complete the testing suite, please only add the label once the PR is code complete and local testing has been performed.

[dsikka]

Do we know why the lm_head in the sparse_2of4_only e2e test was being skipped by the sparse24 bitmask compressor?

[dsikka]

Is this ready for review? still draft

[rahul-tuli]

Is this ready for review? still draft

No it wasn't, now is!