Disable kernels during calibration (and tracing) (#1454)

## Purpose ##
* Guarantee that module hooks trigger by disabling kernel acceleration

## Background ##
* As of `transformers>=4.52`, model forward functions may be overwritten
with kernels
* Kernel execution can be disabled using the config
([source](https://github.com/huggingface/transformers/blob/main/src/transformers/integrations/hub_kernels.py#L84))
* It seems that HF wants to continue enabling regular execution through
this disabling feature. This gives us some faith that they will not
[override the cls init
function](https://github.com/huggingface/transformers/blob/main/src/transformers/integrations/hub_kernels.py#L78)
in a way that inhibits execution with the standard forward definition
* This only affects users who have the hf `kernels` library installed

## Changes ##
* Implement `disable_hf_kernels` context and add to
`calibration_forward_context`
* Remove parenthesis around `calibration_forward_context` with in order
to support python3.9
* Apply style to tests

Signed-off-by: Kyle Sayers <[email protected]>
Co-authored-by: Dipika Sikka <[email protected]>

Author: kylesayrs

src/llmcompressor/utils/helpers.py

@@ -66,6 +66,7 @@
     "eval_context",
     "calibration_forward_context",
     "patch_attr",
+    "disable_hf_kernels",
 ]
 
 
@@ -1024,6 +1025,9 @@ def DisableQuantization(module: torch.nn.Module):
 
 @contextlib.contextmanager
 def eval_context(module: torch.nn.Module):
+    """
+    Disable pytorch training mode for the given module
+    """
     restore_value = module.training
     try:
         module.train(False)  # equivalent to eval()
@@ -1033,6 +1037,21 @@ def eval_context(module: torch.nn.Module):
         module.train(restore_value)
 
 
+@contextlib.contextmanager
+def disable_hf_kernels(model: PreTrainedModel):
+    """
+    In transformers>=4.50.0, some module forward methods may be
+    replaced by calls to hf hub kernels. This has the potential
+    to bypass hooks added by LLM Compressor
+    """
+    if hasattr(model, "config"):
+        with patch_attr(model.config, "disable_custom_kernels", True):
+            yield
+
+    else:
+        yield
+
+
 @contextlib.contextmanager
 def calibration_forward_context(model: PreTrainedModel):
     """
@@ -1041,12 +1060,11 @@ def calibration_forward_context(model: PreTrainedModel):
     - Remove gradient calculations
     - Disable the KV cache
     - Disable train mode and enable eval mode
+    - Disable hf kernels which could bypass hooks
     """
-    with (
-        torch.no_grad(),
-        DisableKVCache(model),
-        eval_context(model),
-    ):
+    with torch.no_grad(), DisableKVCache(model), eval_context(
+        model
+    ), disable_hf_kernels(model):
         yield
 
 

tests/examples/test_quantization_2of4_sparse_w4a16.py

@@ -36,9 +36,9 @@ def test_doc_example_command(self, example_dir: str, tmp_path: Path):
         readme = ReadMe(readme_path)
 
         command = readme.get_code_block_content(position=2, lang="shell")
-        assert command.startswith("python"), (
-            "Expected shell command to start with 'python'"
-        )
+        assert command.startswith(
+            "python"
+        ), "Expected shell command to start with 'python'"
 
         command = shlex.split(command)
         result = copy_and_run_command(tmp_path, example_dir, command)
@@ -62,18 +62,16 @@ def test_doc_example_command(self, example_dir: str, tmp_path: Path):
         }
 
         for stage, stage_info in stages.items():
-            stage_path = (
-                tmp_path / example_dir / output_dir / stage_info["path"]
-            )
+            stage_path = tmp_path / example_dir / output_dir / stage_info["path"]
             recipe_path = stage_path / "recipe.yaml"
             config_path = stage_path / "config.json"
 
-            assert recipe_path.exists(), (
-                f"Missing recipe file in {stage}: {recipe_path}"
-            )
-            assert config_path.exists(), (
-                f"Missing config file in {stage}: {config_path}"
-            )
+            assert (
+                recipe_path.exists()
+            ), f"Missing recipe file in {stage}: {recipe_path}"
+            assert (
+                config_path.exists()
+            ), f"Missing config file in {stage}: {config_path}"
 
             config = AutoConfig.from_pretrained(stage_path)
             assert config is not None, f"Failed to load config in {stage}"
@@ -82,13 +80,9 @@ def test_doc_example_command(self, example_dir: str, tmp_path: Path):
             if stage == "quantization":
                 actual_format = quant_config.get("format")
             else:
-                actual_format = quant_config.get(
-                    "sparsity_config", {}
-                ).get("format")
+                actual_format = quant_config.get("sparsity_config", {}).get("format")
 
-            assert actual_format, (
-                f"Missing expected format field in {stage} config"
-            )
+            assert actual_format, f"Missing expected format field in {stage} config"
             assert actual_format == stage_info["format"], (
                 f"Unexpected format in {stage}: got '{actual_format}', "
                 f"expected '{stage_info['format']}'"