WIP: incorrect offloading

Signed-off-by: Kyle Sayers <[email protected]>

Author: kylesayrs

examples/quantization_w4a16/gpt_oss_example.py

@@ -0,0 +1,79 @@
+from datasets import load_dataset
+from transformers import AutoModelForCausalLM, AutoTokenizer
+
+from llmcompressor.modifiers.quantization import GPTQModifier
+from llmcompressor.transformers import oneshot
+from llmcompressor.utils import dispatch_for_generation
+
+# Select model and load it.
+model_id = "unsloth/gpt-oss-20b-BF16"
+model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype="auto")
+tokenizer = AutoTokenizer.from_pretrained(model_id)
+
+# 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 = 512
+MAX_SEQUENCE_LENGTH = 2048
+
+# Load dataset and preprocess.
+ds = load_dataset(DATASET_ID, split=f"{DATASET_SPLIT}[:{NUM_CALIBRATION_SAMPLES}]")
+ds = ds.shuffle(seed=42)
+
+
+def preprocess(example):
+    return {
+        "text": tokenizer.apply_chat_template(
+            example["messages"],
+            tokenize=False,
+        )
+    }
+
+
+ds = ds.map(preprocess)
+
+
+# Tokenize inputs.
+def tokenize(sample):
+    return tokenizer(
+        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"])
+
+# Apply algorithms.
+oneshot(
+    model=model,
+    dataset=ds,
+    recipe=recipe,
+    max_seq_length=MAX_SEQUENCE_LENGTH,
+    num_calibration_samples=NUM_CALIBRATION_SAMPLES,
+    calibrate_moe_context=True,
+)
+
+# Confirm generations of the quantized model look sane.
+print("\n\n")
+print("========== SAMPLE GENERATION ==============")
+dispatch_for_generation(model)
+sample = tokenizer("Hello my name is", return_tensors="pt")
+sample = {key: value.to("cuda") for key, value in sample.items()}
+output = model.generate(**sample, max_new_tokens=100)
+print(tokenizer.decode(output[0]))
+print("==========================================\n\n")
+
+# Save to disk compressed.
+SAVE_DIR = model_id.rstrip("/").split("/")[-1] + "-W4A16-G128"
+model.save_pretrained(SAVE_DIR, save_compressed=True)
+tokenizer.save_pretrained(SAVE_DIR)

src/llmcompressor/modeling/gpt_oss.py

@@ -7,10 +7,14 @@
 from transformers.models.gpt_oss.configuration_gpt_oss import GptOssConfig
 from llmcompressor.utils.dev import skip_weights_initialize
 
-from compressed_tensors import update_offload_parameter
+from compressed_tensors.utils import update_offload_parameter, align_module_device
 
 
 class GptOssExpert(torch.nn.Module):
+    gate_proj: torch.nn.Linear
+    up_proj: torch.nn.Linear
+    down_proj: torch.nn.Linear
+
     def __init__(self, hidden_size: int, expert_dim: int, alpha: float, limit: float):
         super().__init__()
 
@@ -57,17 +61,21 @@ def __init__(self, experts: GptOssExpert):
         self.limit = experts.limit
 
     def load_weights(self, experts: GptOssExperts):
-        for expert_index, expert in enumerate(self.experts):
-            update_offload_parameter(expert.gate_proj, "weight", experts.gate_up_proj[expert_index, ..., ::2].T)
-            update_offload_parameter(expert.gate_proj, "bias", experts.gate_up_proj_bias[expert_index, ..., ::2])
+        # TODO: this code is inefficient. If there was a "get_offloaded_data" util,
+        # we could avoid having to move from cpu -> gpu -> cpu
+        with align_module_device(experts):
+            for expert_index, expert in enumerate(self.experts):
+                update_offload_parameter(expert.gate_proj, "weight", experts.gate_up_proj[expert_index, ..., ::2].T)
+                update_offload_parameter(expert.gate_proj, "bias", experts.gate_up_proj_bias[expert_index, ..., ::2])
 
-            update_offload_parameter(expert.up_proj, "weight", experts.gate_up_proj[expert_index, ..., 1::2].T)
-            update_offload_parameter(expert.up_proj, "bias", experts.gate_up_proj_bias[expert_index, ..., 1::2])
+                update_offload_parameter(expert.up_proj, "weight", experts.gate_up_proj[expert_index, ..., 1::2].T)
+                update_offload_parameter(expert.up_proj, "bias", experts.gate_up_proj_bias[expert_index, ..., 1::2])
 
-            update_offload_parameter(expert.down_proj, "weight", experts.down_proj[expert_index].T)
-            update_offload_parameter(expert.down_proj, "bias", experts.down_proj_bias[expert_index])
+                update_offload_parameter(expert.down_proj, "weight", experts.down_proj[expert_index].T)
+                update_offload_parameter(expert.down_proj, "bias", experts.down_proj_bias[expert_index])
 
     def to_original(self) -> GptOssExperts:
+        # TODO: this doesn't really handle offloading or correct device placement
         with skip_weights_initialize():
             fake_config = GptOssConfig(
                 intermediate_size=self.intermediate_size,
@@ -78,14 +86,17 @@ def to_original(self) -> GptOssExperts:
             experts = GptOssExperts(fake_config)
 
         for expert_index, expert in enumerate(self.experts):
-            experts.gate_up_proj[expert_index, ..., ::2].data = expert.gate_proj.weight.data.T
-            experts.gate_up_proj_bias[expert_index, ..., ::2].data = expert.gate_proj.bias.data
+            # TODO: this code is inefficient. If there was a "get_offloaded_data" util,
+            # we could avoid having to move from cpu -> gpu -> cpu
+            with align_module_device(expert):
+                experts.gate_up_proj[expert_index, ..., ::2].data = expert.gate_proj.weight.data.T
+                experts.gate_up_proj_bias[expert_index, ..., ::2].data = expert.gate_proj.bias.data
 
-            experts.gate_up_proj[expert_index, ..., 1::2].data = expert.up_proj.weight.data.T
-            experts.gate_up_proj_bias[expert_index, ..., 1::2].data = expert.up_proj.bias.data
+                experts.gate_up_proj[expert_index, ..., 1::2].data = expert.up_proj.weight.data.T
+                experts.gate_up_proj_bias[expert_index, ..., 1::2].data = expert.up_proj.bias.data
 
-            experts.down_proj[expert_index].data = expert.down_proj.weight.data.T
-            experts.down_proj_bias[expert_index] = expert.down_proj.bias.data
+                experts.down_proj[expert_index].data = expert.down_proj.weight.data.T
+                experts.down_proj_bias[expert_index] = expert.down_proj.bias.data
 
         # update offloaded state dict
         update_offload_parameter(experts, "gate_up_proj", experts.gate_up_proj)
@@ -134,6 +145,7 @@ def forward(self, hidden_states: torch.Tensor, router_indices=None, routing_weig
             setattr(original, name, getattr(original, name).normal_())
 
         original.eval()
+        assert original.training == False
         true_output = original(input, routing_weights=routing_weights)
 
         linear = GptOssExpertsLinear(original)

src/llmcompressor/modeling/prepare.py

@@ -1,5 +1,6 @@
+import tqdm
 import contextlib
-from compressed_tensors.utils import replace_module
+from compressed_tensors.utils import replace_module, match_named_modules
 from transformers import PreTrainedModel
 
 from llmcompressor.modeling.deepseek_v3 import replace as replace_deepseekv3
@@ -57,9 +58,10 @@ def replace_context(model, name, module):
         replace_module(model, name, restored)
 
     # TODO: need to think about duplicates
-    for name, module in model.named_modules():
+    modules = list(model.named_modules())
+    for name, module in tqdm.tqdm(modules, desc="Checking modules for replacements"):
         cls_name = module.__class__.__name__
-        if cls_name == "GptOssExpert":
+        if cls_name == "GptOssExperts":
             stack.enter_context(replace_context(model, name, module))