[Llama4] Update MoE Support for LLama4 + Add NVFP4 and W4A16 Examples (#1608)

Summary:

- Updates prepare method to no longer require a replace function but
just pass in the orignal module directly along with the text config
- Add llama4 calibration support - swaps `Llama4TextMoe` with
`SequentialLlama4TextMoe` modules
- Add llama4 example for NVFP4 and W4A16 

Testing
- Tested llama4 NVFP4 e2e to produce:
`nm-testing/Llama-4-Scout-17B-16E-Instruct-NVFP4`

---------

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

Author: kylesayrs

examples/multimodal_vision/llama4_example.py

@@ -0,0 +1,92 @@
+import torch
+from datasets import load_dataset
+from transformers import Llama4ForConditionalGeneration, Llama4Processor
+
+from llmcompressor import oneshot
+from llmcompressor.modeling import prepare_for_calibration
+from llmcompressor.modifiers.quantization import GPTQModifier
+
+# Select model and load it.
+model_id = "meta-llama/Llama-4-Scout-17B-16E-Instruct"
+model = Llama4ForConditionalGeneration.from_pretrained(model_id, torch_dtype="auto")
+processor = Llama4Processor.from_pretrained(model_id)
+# We update `Llama4TextMoe` modules with custom `SequentialLlama4TextMoe`.
+# This change allows compatibility with vllm.
+# To apply your own custom module for experimentation, consider updating
+# `SequentialLlama4TextMoe` under llmcompressor/modeling/llama4.py
+model = prepare_for_calibration(model)
+
+DATASET_ID = "neuralmagic/calibration"
+NUM_CALIBRATION_SAMPLES = 512
+MAX_SEQUENCE_LENGTH = 8192
+
+ds = load_dataset(DATASET_ID, name="LLM", split=f"train[:{NUM_CALIBRATION_SAMPLES}]")
+
+
+def preprocess_function(example):
+    messgages = []
+    for message in example["messages"]:
+        messgages.append(
+            {
+                "role": message["role"],
+                "content": [{"type": "text", "text": message["content"]}],
+            }
+        )
+
+    return processor.apply_chat_template(
+        messgages,
+        return_tensors="pt",
+        padding=False,
+        truncation=True,
+        max_length=MAX_SEQUENCE_LENGTH,
+        tokenize=True,
+        add_special_tokens=False,
+        return_dict=True,
+        add_generation_prompt=False,
+    )
+
+
+ds = ds.map(preprocess_function, batched=False, remove_columns=ds.column_names)
+
+
+def data_collator(batch):
+    assert len(batch) == 1
+    return {
+        key: torch.tensor(value)
+        if key != "pixel_values"
+        else torch.tensor(value, dtype=torch.bfloat16).squeeze(0)
+        for key, value in batch[0].items()
+    }
+
+
+# Configure the quantization algorithm to run.
+recipe = GPTQModifier(
+    targets="Linear",
+    scheme="W4A16",
+    ignore=[
+        "re:.*lm_head",
+        "re:.*self_attn",
+        "re:.*router",
+        "re:vision_model.*",
+        "re:multi_modal_projector.*",
+        "Llama4TextAttention",
+    ],
+)
+
+# Apply algorithms.
+# due to the large size of Llama4, we specify sequential targets such that
+# only one MLP is loaded into GPU memory at a time
+oneshot(
+    model=model,
+    dataset=ds,
+    recipe=recipe,
+    max_seq_length=MAX_SEQUENCE_LENGTH,
+    num_calibration_samples=NUM_CALIBRATION_SAMPLES,
+    data_collator=data_collator,
+    sequential_targets=["Llama4TextMLP"],
+)
+
+# Save to disk compressed.
+SAVE_DIR = model_id.rstrip("/").split("/")[-1] + "-W4A16-G128"
+model.save_pretrained(SAVE_DIR, save_compressed=True)
+processor.save_pretrained(SAVE_DIR)

examples/quantization_w4a4_fp4/llama4_example.py

@@ -0,0 +1,93 @@
+import torch
+from datasets import load_dataset
+from transformers import Llama4ForConditionalGeneration, Llama4Processor
+
+from llmcompressor import oneshot
+from llmcompressor.modeling import prepare_for_calibration
+from llmcompressor.modifiers.quantization import QuantizationModifier
+
+# Select model and load it.
+model_id = "meta-llama/Llama-4-Scout-17B-16E-Instruct"
+model = Llama4ForConditionalGeneration.from_pretrained(model_id, torch_dtype="auto")
+processor = Llama4Processor.from_pretrained(model_id)
+# We update `Llama4TextMoe` modules with custom `SequentialLlama4TextMoe`.
+# This change allows compatibility with vllm.
+# To apply your own custom module for experimentation, consider updating
+# `SequentialLlama4TextMoe` under llmcompressor/modeling/llama4.py
+model = prepare_for_calibration(model)
+
+DATASET_ID = "neuralmagic/calibration"
+NUM_CALIBRATION_SAMPLES = 20
+MAX_SEQUENCE_LENGTH = 8192
+
+ds = load_dataset(DATASET_ID, name="LLM", split=f"train[:{NUM_CALIBRATION_SAMPLES}]")
+
+
+def preprocess_function(example):
+    messgages = []
+    for message in example["messages"]:
+        messgages.append(
+            {
+                "role": message["role"],
+                "content": [{"type": "text", "text": message["content"]}],
+            }
+        )
+
+    return processor.apply_chat_template(
+        messgages,
+        return_tensors="pt",
+        padding=False,
+        truncation=True,
+        max_length=MAX_SEQUENCE_LENGTH,
+        tokenize=True,
+        add_special_tokens=False,
+        return_dict=True,
+        add_generation_prompt=False,
+    )
+
+
+ds = ds.map(preprocess_function, batched=False, remove_columns=ds.column_names)
+
+
+def data_collator(batch):
+    assert len(batch) == 1
+    return {
+        key: torch.tensor(value)
+        if key != "pixel_values"
+        else torch.tensor(value, dtype=torch.bfloat16).squeeze(0)
+        for key, value in batch[0].items()
+    }
+
+
+# Configure the quantization algorithm to run.
+recipe = QuantizationModifier(
+    targets="Linear",
+    scheme="NVFP4",
+    ignore=[
+        "re:.*lm_head",
+        "re:.*self_attn",
+        "re:.*router",
+        "re:vision_model.*",
+        "re:multi_modal_projector.*",
+        "Llama4TextAttention",
+    ],
+)
+
+# Apply algorithms.
+# due to the large size of Llama4, we specify sequential targets such that
+# only one MLP is loaded into GPU memory at a time
+oneshot(
+    model=model,
+    dataset=ds,
+    recipe=recipe,
+    max_seq_length=MAX_SEQUENCE_LENGTH,
+    num_calibration_samples=NUM_CALIBRATION_SAMPLES,
+    sequential_targets=["Llama4TextMLP"],
+    data_collator=data_collator,
+)
+
+
+# Save to disk compressed.
+SAVE_DIR = model_id.rstrip("/").split("/")[-1] + "-NVFP4"
+model.save_pretrained(SAVE_DIR)
+processor.save_pretrained(SAVE_DIR)

src/llmcompressor/modeling/deepseek_v3.py

@@ -1,20 +1,23 @@
 import torch
+from transformers.models.deepseek_v3.configuration_deepseek_v3 import DeepseekV3Config
 from transformers.models.deepseek_v3.modeling_deepseek_v3 import DeepseekV3MoE
 
+__all__ = ["DeepseekV3MoECalibrate"]
+
 
 class DeepseekV3MoECalibrate(torch.nn.Module):
     """
     Patched DeepseekV3MoE which sends all tokens to all experts for calibration
     """
 
-    def __init__(self, config, experts, gate, shared_experts):
+    def __init__(self, config: DeepseekV3Config, original: DeepseekV3MoE):
         super().__init__()
         self.config = config
-        self.experts = experts
-        self.gate = gate
-        self.shared_experts = shared_experts
+        self.experts = original.experts
+        self.gate = original.gate
+        self.shared_experts = original.shared_experts
 
-    def forward(self, hidden_states):
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
         residuals = hidden_states
         orig_shape = hidden_states.shape
         topk_indices, topk_weights = self.gate(hidden_states)
@@ -46,7 +49,5 @@ def forward(self, hidden_states):
         return hidden_states
 
 
-def replace(module: DeepseekV3MoE) -> DeepseekV3MoECalibrate:
-    return DeepseekV3MoECalibrate(
-        module.config, module.experts, module.gate, module.shared_experts
-    )
+def replace(config: DeepseekV3Config, module: DeepseekV3MoE):
+    return DeepseekV3MoECalibrate(config=config, original=module)

src/llmcompressor/modeling/llama4.py

@@ -0,0 +1,68 @@
+from typing import Tuple
+
+import torch
+from transformers.models import Llama4Config
+from transformers.models.llama4.configuration_llama4 import Llama4TextConfig
+from transformers.models.llama4.modeling_llama4 import (
+    Llama4TextExperts,
+    Llama4TextMLP,
+    Llama4TextMoe,
+)
+
+from llmcompressor.utils.dev import skip_weights_initialize
+
+__all__ = ["SequentialLlama4TextMoe"]
+
+
+class SequentialLlama4TextMoe(torch.nn.Module):
+    def __init__(self, config: Llama4TextConfig, original: Llama4TextMoe):
+        super().__init__()
+        self.top_k = config.num_experts_per_tok
+        self.hidden_dim = config.hidden_size
+        self.num_experts = config.num_local_experts
+        self.experts = SequentialLlama4TextExperts(config, original.experts)
+        self.router = original.router
+        self.shared_expert = original.shared_expert
+
+    def forward(self, hidden_states: torch.Tensor) -> Tuple[torch.Tensor, torch.tensor]:
+        hidden_states = hidden_states.reshape(-1, self.hidden_dim)
+        router_logits = self.router(hidden_states)
+
+        router_top_value, router_indices = torch.topk(router_logits, self.top_k, dim=1)
+
+        router_scores = (
+            torch.full_like(router_logits, float("-inf"))
+            .scatter_(1, router_indices, router_top_value)
+            .transpose(0, 1)
+        )
+        router_scores = torch.sigmoid(router_scores.float()).to(hidden_states.dtype)
+
+        out = self.shared_expert(hidden_states)
+        for i in range(self.num_experts):
+            out += self.experts[i](hidden_states) * router_scores[i].reshape(-1, 1)
+
+        return out, router_scores
+
+
+class SequentialLlama4TextExperts(torch.nn.ModuleList):
+    def __init__(self, config: Llama4TextConfig, original: Llama4TextExperts):
+        self.num_experts = original.gate_up_proj.shape[0]
+        with skip_weights_initialize():
+            super().__init__([Llama4TextMLP(config) for _ in range(self.num_experts)])
+
+        intermediate_size = original.down_proj.shape[1]
+
+        for i in range(self.num_experts):
+            gate_up = original.gate_up_proj[i]
+            down = original.down_proj[i]
+
+            gate_proj = gate_up[:, :intermediate_size]
+            up_proj = gate_up[:, intermediate_size:]
+
+            self[i].gate_proj.weight.data = gate_proj.t().clone().contiguous()
+            self[i].up_proj.weight.data = up_proj.t().clone().contiguous()
+            self[i].down_proj.weight.data = down.t().clone().contiguous()
+
+
+def replace(config: Llama4Config, module: Llama4TextMoe):
+    return SequentialLlama4TextMoe(config=config.get_text_config(), original=module)

src/llmcompressor/modeling/prepare.py

@@ -1,20 +1,22 @@
 from compressed_tensors.utils import replace_module
 from transformers import PreTrainedModel
 
-from llmcompressor.modeling.deepseek_v3 import replace as replace_DeepseekV3MoE
+from llmcompressor.modeling.deepseek_v3 import replace as replace_deepseekv3
+from llmcompressor.modeling.llama4 import replace as replace_llama4
 
 __all__ = ["prepare_for_calibration"]
 
 replacements = {
-    "DeepseekV3MoE": replace_DeepseekV3MoE,
+    "DeepseekV3MoE": replace_deepseekv3,
+    "Llama4TextMoe": replace_llama4,
 }
 
 
 def prepare_for_calibration(model: PreTrainedModel) -> PreTrainedModel:
     for name, module in model.named_modules():
         cls_name = module.__class__.__name__
         if cls_name in replacements:
-            new_module = replacements[cls_name](module)
+            new_module = replacements[cls_name](config=model.config, module=module)
             replace_module(model, name, new_module)
 
     return model