Unpin to support transformers==4.52.3 (#1479)

## Purpose ##
* Support the latest transformers release

## Prerequisites ##
* #1481
* #1411

## Fixes ##
* #1457

## Changes ##
* Unpin transformers version
* Add `torchvision`, `librosa`, and `soundfile` to dev dependencies
(needed to test models)
* Fix default ignore list for tracing debugger
* Add back llama4 model tests

---------

Signed-off-by: Kyle Sayers <kylesayrs@gmail.com>

Author: kylesayrs

setup.py

@@ -116,8 +116,7 @@ def localversion_func(version: ScmVersion) -> str:
         "requests>=2.0.0",
         "tqdm>=4.0.0",
         "torch>=1.7.0",
-        # TODO (#1457) revert to "transformers>4.0,<5.0" after tests pass
-        "transformers>4.0,<4.52.0",
+        "transformers>4.0,<5.0",
         "datasets",
         "accelerate>=0.20.3,!=1.1.0",
         "pynvml",
@@ -136,11 +135,14 @@ def localversion_func(version: ScmVersion) -> str:
             "pytest-rerunfailures>=13.0",
             "parameterized",
             "lm_eval==0.4.5",
-            # example test dependencies
+            # test dependencies
             "beautifulsoup4~=4.12.3",
             "cmarkgfm~=2024.1.14",
             "trl>=0.10.1",
             "pandas",
+            "torchvision",
+            "librosa",
+            "soundfile",
             # linting, formatting, and type checking
             "black~=24.4.2",
             "isort~=5.13.2",

src/llmcompressor/transformers/tracing/debug.py

@@ -23,7 +23,7 @@ def parse_args():
     parser.add_argument("--model_id", type=str, required=True, help="The stub of the model to load")  # noqa: E501
     parser.add_argument("--model_class", type=str, required=True, help="The class name of the model")  # noqa: E501
     parser.add_argument("--sequential_targets", type=str, nargs="*", default=None, metavar="TARGET", help="List of targets for sequential tracing")  # noqa: E501
-    parser.add_argument("--ignore", type=str, nargs="*", default=[], metavar="PATTERN", help="List of patterns to ignore during tracing")  # noqa: E501
+    parser.add_argument("--ignore", type=str, nargs="*", default=["_update_causal_mask"], metavar="PATTERN", help="List of patterns to ignore during tracing")  # noqa: E501
     parser.add_argument("--modality", type=str, default="text", help="Modality of calibration dataset, defaults to text")  # noqa: E501
     parser.add_argument("--trust_remote_code", type=bool, default=False, help="Whether to trust model remote code")  # noqa: E501
     parser.add_argument("--skip_weights", type=bool, default=True, help="Whether to load the model with dummy weights")  # noqa: E501

tests/llmcompressor/transformers/tracing/test_models.py

@@ -3,6 +3,7 @@
     AutoModelForCausalLM,
     Gemma3ForConditionalGeneration,
     Idefics3ForConditionalGeneration,
+    Llama4ForConditionalGeneration,
     LlavaForConditionalGeneration,
     MllamaForConditionalGeneration,
     Qwen2_5_VLForConditionalGeneration,
@@ -86,21 +87,20 @@
             "vision",
             [],
         ),
-        # TODO: add back once transformers is bumped
-        # (
-        #     "meta-llama/Llama-4-Scout-17B-16E-Instruct",
-        #     Llama4ForConditionalGeneration,
-        #     "Llama4TextDecoderLayer",
-        #     "vision",
-        #     [],
-        # ),
-        # (
-        #     "meta-llama/Llama-4-Maverick-17B-128E-Instruct",
-        #     Llama4ForConditionalGeneration,
-        #     "Llama4TextDecoderLayer",
-        #     "vision",
-        #     [],
-        # ),
+        (
+            "meta-llama/Llama-4-Scout-17B-16E-Instruct",
+            Llama4ForConditionalGeneration,
+            "Llama4TextDecoderLayer",
+            "vision",
+            [],
+        ),
+        (
+            "meta-llama/Llama-4-Maverick-17B-128E-Instruct",
+            Llama4ForConditionalGeneration,
+            "Llama4TextDecoderLayer",
+            "vision",
+            [],
+        ),
         # --- audio ---
         (
             "openai/whisper-large-v3",