support vision datasets

examples/quantization_w4a16/llama3_example.py

@@ -9,7 +9,7 @@
 
 model = SparseAutoModelForCausalLM.from_pretrained(
     MODEL_ID,
-    device_map="auto",
+    device_map="cuda:0",
     torch_dtype="auto",
 )
 tokenizer = AutoTokenizer.from_pretrained(MODEL_ID)
@@ -20,7 +20,7 @@
 
 # Select number of samples. 512 samples is a good place to start.
 # Increasing the number of samples can improve accuracy.
-NUM_CALIBRATION_SAMPLES = 512
+NUM_CALIBRATION_SAMPLES = 160 #2048
 MAX_SEQUENCE_LENGTH = 2048
 
 # Load dataset and preprocess.
@@ -55,7 +55,7 @@ def tokenize(sample):
 
 # Configure the quantization algorithm to run.
 #   * quantize the weights to 4 bit with GPTQ with a group size 128
-recipe = GPTQModifier(targets="Linear", scheme="W4A16", ignore=["lm_head"])
+recipe = GPTQModifier(targets="Linear", scheme="W4A16", ignore=["lm_head"], batch_size=-1, dampening_frac=0.5)
 
 # Apply algorithms.
 oneshot(

examples/quantization_w4a16/vision2_example.py

@@ -0,0 +1,83 @@
+from datasets import load_dataset
+from transformers import AutoProcessor, MllamaForConditionalGeneration
+
+from llmcompressor.modifiers.quantization import GPTQModifier
+from llmcompressor.transformers import oneshot
+
+# Select model and load it.
+MODEL_ID = "meta-llama/Llama-3.2-11B-Vision-Instruct"
+
+model = MllamaForConditionalGeneration.from_pretrained(
+    MODEL_ID,
+    device_map="cuda:0",
+    torch_dtype="auto",
+)
+processor = AutoProcessor.from_pretrained(MODEL_ID, trust_remote_code=True)
+
+# Select calibration dataset.
+DATASET_ID = "HuggingFaceH4/ultrachat_200k"
+DATASET_SPLIT = "train_sft"
+
+# Select number of samples. 512 samples is a good place to start.
+# Increasing the number of samples can improve accuracy.
+NUM_CALIBRATION_SAMPLES = 160 #2048
+MAX_SEQUENCE_LENGTH = 2048
+
+# Load dataset and preprocess.
+ds = load_dataset(DATASET_ID, split=DATASET_SPLIT)
+ds = ds.shuffle(seed=42).select(range(NUM_CALIBRATION_SAMPLES))
+
+
+def preprocess(example):
+    return {
+        "text": processor.apply_chat_template(
+            example["messages"],
+            tokenize=False,
+        )
+    }
+
+
+ds = ds.map(preprocess)
+
+
+# Tokenize inputs.
+def tokenize(sample):
+    return processor(
+        None,
+        sample["text"],
+        padding=False,
+        max_length=MAX_SEQUENCE_LENGTH,
+        truncation=True,
+        add_special_tokens=False,
+    )
+
+
+ds = ds.map(tokenize, remove_columns=ds.column_names)
+
+# Configure the quantization algorithm to run.
+#   * quantize the weights to 4 bit with GPTQ with a group size 128
+recipe = GPTQModifier(targets="Linear", scheme="W4A16", ignore=["lm_head"], batch_size=1, dampening_frac=0.5)
+
+# Apply algorithms.
+oneshot(
+    model=model,
+    tokenizer=MODEL_ID,
+    dataset=ds,
+    recipe=recipe,
+    max_seq_length=MAX_SEQUENCE_LENGTH,
+    num_calibration_samples=NUM_CALIBRATION_SAMPLES,
+    trust_remote_code_model=True,
+)
+
+# Confirm generations of the quantized model look sane.
+print("\n\n")
+print("========== SAMPLE GENERATION ==============")
+input_ids = processor("Hello my name is", return_tensors="pt").input_ids.to("cuda")
+output = model.generate(input_ids, max_new_tokens=100)
+print(processor.decode(output[0]))
+print("==========================================\n\n")
+
+# 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/quantization_w4a16/vision_example.py

@@ -0,0 +1,88 @@
+from datasets import load_dataset
+from transformers import AutoProcessor
+
+from llmcompressor.modifiers.quantization import GPTQModifier
+from llmcompressor.transformers import SparseAutoModelForCausalLM, oneshot
+
+# Select model and load it.
+MODEL_ID = "meta-llama/Llama-3.2-11B-Vision-Instruct"
+
+model = SparseAutoModelForCausalLM.from_pretrained(
+    MODEL_ID,
+    device_map="cuda:0",
+    torch_dtype="auto",
+)
+breakpoint()
+processor = AutoProcessor.from_pretrained(MODEL_ID, trust_remote_code=True)
+
+# Select calibration dataset.
+DATASET_ID = "lmms-lab/flickr30k"
+DATASET_SPLIT = "test[:165]"
+
+# Select number of samples. 512 samples is a good place to start.
+# Increasing the number of samples can improve accuracy.
+NUM_CALIBRATION_SAMPLES = 165 #2048
+MAX_SEQUENCE_LENGTH = 2048
+
+# Load dataset and preprocess.
+ds = load_dataset(DATASET_ID, split=DATASET_SPLIT)
+ds = ds.shuffle(seed=42).select(range(NUM_CALIBRATION_SAMPLES))
+
+
+def preprocess(example):
+    messages = [
+        [
+            {
+                "role": "user", 
+                "content": [
+                    {"type": "image"},
+                    {"type": "text", "text": "What does the image show?"}
+                ]
+            }
+        ],
+    ]
+    return {
+        "text": processor.apply_chat_template(
+            messages,
+            add_generation_prompt=True,
+        ),
+    }
+
+
+ds = ds.map(preprocess)
+
+
+# Tokenize inputs.
+def tokenize(sample):
+    return processor(sample["image"], sample["text"], add_special_tokens=False, return_tensors="pt", max_length=MAX_SEQUENCE_LENGTH)
+
+
+ds = ds.map(tokenize, remove_columns=ds.column_names)
+
+# Configure the quantization algorithm to run.
+#   * quantize the weights to 4 bit with GPTQ with a group size 128
+recipe = GPTQModifier(targets="Linear", scheme="W4A16", ignore=["lm_head"], batch_size=-1, dampening_frac=0.5)
+
+# Apply algorithms.
+oneshot(
+    model=model,
+    tokenizer=MODEL_ID,
+    dataset=ds,
+    recipe=recipe,
+    max_seq_length=MAX_SEQUENCE_LENGTH,
+    num_calibration_samples=NUM_CALIBRATION_SAMPLES,
+    trust_remote_code_model=True,
+)
+
+# Confirm generations of the quantized model look sane.
+print("\n\n")
+print("========== SAMPLE GENERATION ==============")
+input_ids = processor("Hello my name is", return_tensors="pt").input_ids.to("cuda")
+output = model.generate(input_ids, max_new_tokens=100)
+print(processor.decode(output[0]))
+print("==========================================\n\n")
+
+# 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)

src/llmcompressor/modifiers/quantization/gptq/base.py

@@ -209,7 +209,7 @@ def on_initialize(self, state: "State", **kwargs) -> bool:
             self.calibration_forward(state.model, state.data.calib)
         finally:
             pass
-            #torch.cuda.memory._dump_snapshot("bs256.pickle")
+            #torch.cuda.memory._dump_snapshot("bs10.pickle")
             #torch.cuda.memory._record_memory_history(enabled=None)
             #exit(0)
 
@@ -306,6 +306,7 @@ def compress_module(
         logger.info(f"Using {inp.size(0)} samples")
 
         with align_module(module):
+            print(inp.shape)
             loss, quantized_weight, _scale, _zero_point, _g_idx = quantize_weight(
                 module.weight.data,
                 inp,

src/llmcompressor/modifiers/utils/hooks.py

@@ -124,11 +124,11 @@ def register_hooks(self, model: torch.nn.Module):
         # if no targets are provided, default to the modules that shouldn't be
         # split by FSDP. For Transformers models this is equivalent to the
         # decoder layers (ie LlamaDecoderLayer)
-        sequential_targets = self.sequential_targets
-        if sequential_targets is None:
-            sequential_targets = get_no_split_params(model)
-        layers = get_layers(sequential_targets, model)
-        self._num_layers = len(layers)
+        # sequential_targets = self.sequential_targets
+        # if sequential_targets is None:
+        #     sequential_targets = get_no_split_params(model)
+        # layers = get_layers(sequential_targets, model)
+        # self._num_layers = len(layers)
 
         for name, module in model.named_modules():
             if getattr_chain(module, "quantization_scheme.weights", None) is not None:
@@ -139,13 +139,13 @@ def register_hooks(self, model: torch.nn.Module):
 
                 self.pre_compress_module(module)
 
-            if name in layers.keys():
-                pre_hook = partial(self.layer_pre_forward, name)
-                post_hook = partial(self.layer_post_forward, name)
-                self.register_hook(module.register_forward_pre_hook(pre_hook, with_kwargs=True))
-                self.register_hook(
-                    module.register_forward_hook(post_hook, with_kwargs=True)
-                )
+            # if name in layers.keys():
+            #     pre_hook = partial(self.layer_pre_forward, name)
+            #     post_hook = partial(self.layer_post_forward, name)
+            #     self.register_hook(module.register_forward_pre_hook(pre_hook, with_kwargs=True))
+            #     self.register_hook(
+            #         module.register_forward_hook(post_hook, with_kwargs=True)
+            #     )
 
 
     @HooksMixin.hook

src/llmcompressor/transformers/finetune/data/base.py

@@ -51,17 +51,20 @@ def __init__(
             self.padding = False
 
         if self.tokenizer:
+            if hasattr(self.tokenizer, "tokenizer"):
+                self.tokenizer = self.tokenizer.tokenizer
+
             if not self.tokenizer.pad_token:
                 self.tokenizer.pad_token = self.tokenizer.eos_token
 
         # configure sequence length
         max_seq_length = data_args.max_seq_length
-        model_max_length = tokenizer.model_max_length if tokenizer else max_seq_length
+        model_max_length = self.tokenizer.model_max_length if self.tokenizer else max_seq_length
         if self.tokenizer and max_seq_length > model_max_length:
             logger.warning(
                 f"The max_seq_length passed ({max_seq_length}) is larger than "
-                f"the maximum length for the model ({tokenizer.model_max_length}). "
-                f"Using max_seq_length={tokenizer.model_max_length}."
+                f"the maximum length for the model ({self.tokenizer.model_max_length}). "
+                f"Using max_seq_length={self.tokenizer.model_max_length}."
             )
         self.max_seq_length = min(data_args.max_seq_length, model_max_length)
 

src/llmcompressor/transformers/finetune/data/data_helpers.py

@@ -38,8 +38,8 @@ def create_batch_dataloader(
 
     def pad_sequences(batch):
         # extract input_ids and attention_mask from the batch
-        input_ids = [torch.tensor(item["input_ids"]) for item in batch]
-        masks = [torch.tensor(item["attention_mask"]) for item in batch]
+        input_ids = [torch.tensor(item["input_ids"]).squeeze(0) for item in batch]
+        masks = [torch.tensor(item["attention_mask"]).squeeze(0) for item in batch]
 
         # while 0 is not necessarily the "correct" padding value, the padded
         # input_ids are ignored according to the attention_mask

src/llmcompressor/transformers/finetune/text_generation.py

@@ -23,7 +23,7 @@
 from loguru import logger
 from transformers import (
     AutoConfig,
-    AutoTokenizer,
+    AutoProcessor,
     DefaultDataCollator,
     HfArgumentParser,
     set_seed,
@@ -221,12 +221,12 @@ def initialize_model_from_path(
 def initialize_tokenizer_from_path(model_args, model, teacher):
     tokenizer_src = model_args.tokenizer
     tokenizer_src = tokenizer_src or get_shared_tokenizer_src(model, teacher)
-    tokenizer = AutoTokenizer.from_pretrained(
+    tokenizer = AutoProcessor.from_pretrained(
         tokenizer_src,
-        cache_dir=model_args.cache_dir,
-        use_fast=True,
-        revision=model_args.model_revision,
-        use_auth_token=True if model_args.use_auth_token else None,
+        # cache_dir=model_args.cache_dir,
+        # use_fast=True,
+        # revision=model_args.model_revision,
+        # use_auth_token=True if model_args.use_auth_token else None,
         trust_remote_code=model_args.trust_remote_code_model,
     )