Delete deprecated GemlitePackedLayout and GemliteUIntXWeightOnlyConfig

[jerryzh168]

Stack from ghstack (oldest at bottom):

• Delete deprecated TensorCoreTiledLayout and related code #4153
• Move bitpacking.py to prototype and add uintx_utils.py #4152
• Delete deprecated PlainLayout, PlainAQTTensorImpl and related v1 code paths #4151
• Delete deprecated PackedLinearInt8DynamicActivationIntxWeightLayout and related code #4150
• Delete deprecated Int4CPULayout and related code #4149
• Delete deprecated SemiSparseLayout and related code #4148
• Delete deprecated Int4XPULayout and related code #4147
• Delete deprecated QDQLayout and related code #4146
• Delete deprecated UintxLayout and related code #4145
• -> Delete deprecated GemlitePackedLayout and GemliteUIntXWeightOnlyConfig #4144

Deprecation Notes:
GemliteUIntXWeightOnlyConfig is now replaced with torchao.prototype.UIntxWeightOnlyConfig and Int8DynamicActivationUIntxWeightConfig

# 0.16:
GemliteUIntXWeightOnlyConfig(group_size, bit_width, packing_bitwidth, mode="weight_only")
GemliteUIntXWeightOnlyConfig(group_size, bit_width, packing_bitwidth, mode="dynamic")
# 0.17
UIntxWeightOnlyConfig(group_size, bit_width, packing_bitwidth)
Int8DynamicActivationUIntxWeightConfig(group_size, bit_width, packing_bitwidth)

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/ao/4144

• 📄 Preview Python docs built from this PR

Note: Links to docs will display an error until the docs builds have been completed.

This comment was automatically generated by Dr. CI and updates every 15 minutes.

[jerryzh168]

@claude review

[claude]

Claude finished @jerryzh168's task in 2m 9s —— View job

---

Review: Delete deprecated GemlitePackedLayout and GemliteUIntXWeightOnlyConfig

• Gather context and read the diff
• Review deleted files for completeness
• Review modified files for correctness
• Check for any remaining references to deleted code
• Post review feedback

---

Overall this is a clean deprecation removal. The diff is straightforward — removing the deprecated GemliteUIntXWeightOnlyConfig, GemlitePackedLayout, and all associated code (layout implementation, dispatch registration, tests). The replacement APIs (UIntxWeightOnlyConfig and Int8DynamicActivationUIntxWeightConfig) in torchao.prototype remain intact.

Findings

Leftover references to deleted code (not cleaned up in this PR):

1. benchmarks/microbenchmarks/utils.py:19,244 — Still imports and uses GemliteUIntXWeightOnlyConfig. This will cause an ImportError at runtime when the benchmarking utils are used with the gemlitewo config string.

   Fix this →

2. docs/source/contributing/benchmarking_api_guide.md:34 — Still references GemliteUIntXWeightOnlyConfig as an example in the docs.

   Fix this →

3. torchao/testing/utils.py:180 — The skip_if_no_gemlite() test decorator is still defined. While it's not necessarily wrong to keep it (gemlite is still used by the prototype UIntxWeightOnlyConfig), this PR removes the only test file that imported it (test_affine_quantized.py removed its import). Worth checking if any other tests still use it.

Minor observation in the diff:

4. In test/quantization/test_quant_api.py, the removal of the try: import gemlite block also moved the MXDynamicActivationMXWeightConfig / NVFP4DynamicActivationNVFP4WeightConfig imports to a different position in the file (from after the gemlite block to before the torchao.quantization imports). This is fine — just import reordering with no functional change.

5. In test/quantization/test_quant_api.py:362, after the removal of the two GemliteUIntXWeightOnlyConfig conditionals, there's a dangling blank line left behind (the elif and the dtype override are removed, leaving an extra blank line before dtype = torch.bfloat16). Very minor nit, no functional impact.

Verdict

The core deletion is correct and complete. The two leftover references in benchmarks/ and docs/ (items 1 and 2) should be addressed to avoid import errors and stale documentation. These could be fixed in this PR or a follow-up.

---