VLM Support via GPTQ Hooks and Data Pipelines

examples/multimodal_vision/llava_example.py

@@ -0,0 +1,54 @@
+from transformers import AutoProcessor
+
+from llmcompressor.modifiers.quantization import GPTQModifier
+from llmcompressor.transformers import oneshot
+from llmcompressor.transformers.tracing import TraceableLlavaForConditionalGeneration
+from llmcompressor.transformers.utils.data_collator import llava_data_collator
+
+# Load model.
+model_id = "llava-hf/llava-1.5-7b-hf"
+model = TraceableLlavaForConditionalGeneration.from_pretrained(
+    model_id, device_map="auto", torch_dtype="auto"
+)
+processor = AutoProcessor.from_pretrained(model_id, trust_remote_code=True)
+
+# Oneshot arguments
+DATASET_ID = "flickr30k"
+DATASET_SPLIT = {"calibration": "test[:512]"}
+NUM_CALIBRATION_SAMPLES = 512
+MAX_SEQUENCE_LENGTH = 2048
+
+# Recipe
+recipe = [
+    GPTQModifier(
+        targets="Linear",
+        scheme="W4A16",
+        ignore=["re:.*lm_head", "re:vision_tower.*", "re:multi_modal_projector.*"],
+        sequential_targets=["LlamaDecoderLayer"],
+    ),
+]
+
+# Perform oneshot
+oneshot(
+    model=model,
+    tokenizer=model_id,
+    dataset=DATASET_ID,
+    splits=DATASET_SPLIT,
+    recipe=recipe,
+    max_seq_length=MAX_SEQUENCE_LENGTH,
+    num_calibration_samples=NUM_CALIBRATION_SAMPLES,
+    trust_remote_code_model=True,
+    data_collator=llava_data_collator,
+)
+
+# Confirm generations of the quantized model look sane.
+print("========== SAMPLE GENERATION ==============")
+input_ids = processor(text="Hello my name is", return_tensors="pt").input_ids.to("cuda")
+output = model.generate(input_ids, max_new_tokens=20)
+print(processor.decode(output[0]))
+print("==========================================")
+
+# Save to disk compressed.
+SAVE_DIR = model_id.split("/")[1] + "-W4A16-G128"
+model.save_pretrained(SAVE_DIR, save_compressed=True)
+processor.save_pretrained(SAVE_DIR)

examples/multimodal_vision/mllama_example.py

@@ -0,0 +1,53 @@
+from transformers import AutoProcessor
+
+from llmcompressor.modifiers.quantization import GPTQModifier
+from llmcompressor.transformers import oneshot
+from llmcompressor.transformers.tracing import TraceableMllamaForConditionalGeneration
+from llmcompressor.transformers.utils.data_collator import mllama_data_collator
+
+# Load model.
+model_id = "meta-llama/Llama-3.2-11B-Vision-Instruct"
+model = TraceableMllamaForConditionalGeneration.from_pretrained(
+    model_id, device_map="auto", torch_dtype="auto"
+)
+processor = AutoProcessor.from_pretrained(model_id, trust_remote_code=True)
+
+# Oneshot arguments
+DATASET_ID = "flickr30k"
+DATASET_SPLIT = {"calibration": "test[:512]"}
+NUM_CALIBRATION_SAMPLES = 512
+MAX_SEQUENCE_LENGTH = 2048
+
+# Recipe
+recipe = [
+    GPTQModifier(
+        targets="Linear",
+        scheme="W4A16",
+        ignore=["re:.*lm_head", "re:multi_modal_projector.*", "re:vision_model.*"],
+    ),
+]
+
+# Perform oneshot
+oneshot(
+    model=model,
+    tokenizer=model_id,
+    dataset=DATASET_ID,
+    splits=DATASET_SPLIT,
+    recipe=recipe,
+    max_seq_length=MAX_SEQUENCE_LENGTH,
+    num_calibration_samples=NUM_CALIBRATION_SAMPLES,
+    trust_remote_code_model=True,
+    data_collator=mllama_data_collator,
+)
+
+# Confirm generations of the quantized model look sane.
+print("========== SAMPLE GENERATION ==============")
+input_ids = processor(text="Hello my name is", return_tensors="pt").input_ids.to("cuda")
+output = model.generate(input_ids, max_new_tokens=20)
+print(processor.decode(output[0]))
+print("==========================================")
+
+# Save to disk compressed.
+SAVE_DIR = model_id.split("/")[1] + "-W4A16-G128"
+model.save_pretrained(SAVE_DIR, save_compressed=True)
+processor.save_pretrained(SAVE_DIR)

examples/multimodal_vision/pixtral_example.py

@@ -0,0 +1,54 @@
+from transformers import AutoProcessor
+
+from llmcompressor.modifiers.quantization import GPTQModifier
+from llmcompressor.transformers import oneshot
+from llmcompressor.transformers.tracing import TraceableLlavaForConditionalGeneration
+from llmcompressor.transformers.utils.data_collator import pixtral_data_collator
+
+# Load model.
+model_id = "mgoin/pixtral-12b"
+model = TraceableLlavaForConditionalGeneration.from_pretrained(
+    model_id, device_map="auto", torch_dtype="auto"
+)
+processor = AutoProcessor.from_pretrained(model_id, trust_remote_code=True)
+
+# Oneshot arguments
+DATASET_ID = "flickr30k"
+DATASET_SPLIT = {"calibration": "test[:512]"}
+NUM_CALIBRATION_SAMPLES = 512
+MAX_SEQUENCE_LENGTH = 2048
+
+# Recipe
+recipe = [
+    GPTQModifier(
+        targets="Linear",
+        scheme="W4A16",
+        ignore=["re:.*lm_head", "re:vision_tower.*", "re:multi_modal_projector.*"],
+        sequential_targets=["MistralDecoderLayer"],
+    ),
+]
+
+# Perform oneshot
+oneshot(
+    model=model,
+    tokenizer=model_id,
+    dataset=DATASET_ID,
+    splits=DATASET_SPLIT,
+    recipe=recipe,
+    max_seq_length=MAX_SEQUENCE_LENGTH,
+    num_calibration_samples=NUM_CALIBRATION_SAMPLES,
+    trust_remote_code_model=True,
+    data_collator=pixtral_data_collator,
+)
+
+# Confirm generations of the quantized model look sane.
+print("========== SAMPLE GENERATION ==============")
+input_ids = processor(text="Hello my name is", return_tensors="pt").input_ids.to("cuda")
+output = model.generate(input_ids, max_new_tokens=20)
+print(processor.decode(output[0]))
+print("==========================================")
+
+# Save to disk compressed.
+SAVE_DIR = model_id.split("/")[1] + "-W4A16-G128"
+model.save_pretrained(SAVE_DIR, save_compressed=True)
+processor.save_pretrained(SAVE_DIR)

pyproject.toml

@@ -8,12 +8,13 @@ target-version = ['py38']
 
 [tool.isort]
 profile = "black"
+skip = ["src/llmcompressor/transformers/tracing/"]
 
 [tool.mypy]
 files = "src/guidellm"
 
 [tool.ruff]
-exclude = ["build", "dist", "env", ".venv"]
+exclude = ["build", "dist", "env", ".venv", "src/llmcompressor/transformers/tracing/"]
 lint.select = ["E", "F", "W"]
 
 [tool.flake8]