feat: add Qwen3.5 MoE calibration module

examples/quantization_w4a4_fp4/qwen3_5_moe_example.py

@@ -0,0 +1,140 @@
+import os
+import shutil
+
+from datasets import concatenate_datasets, load_dataset
+from huggingface_hub import snapshot_download
+from transformers import AutoModelForImageTextToText, AutoTokenizer
+
+from llmcompressor import oneshot
+from llmcompressor.modifiers.quantization import QuantizationModifier
+from llmcompressor.utils import dispatch_for_generation
+
+# Available Qwen3.5 MoE models (pick one):
+#   "Qwen/Qwen3.5-35B-A3B"
+#   "Qwen/Qwen3.5-122B-A10B"
+#   "Qwen/Qwen3.5-397B-A17B"
+MODEL_ID = "Qwen/Qwen3.5-35B-A3B"
+
+# Load model.
+model = AutoModelForImageTextToText.from_pretrained(MODEL_ID, dtype="auto")
+tokenizer = AutoTokenizer.from_pretrained(MODEL_ID)
+
+# Select number of samples. 512 is recommended for production quality;
+# reduce to 256 or lower for faster iteration during development.
+NUM_CALIBRATION_SAMPLES = 256
+MAX_SEQUENCE_LENGTH = 4096
+
+# Load datasets and preprocess.
+# Use half from each source for a diverse calibration set.
+samples_per_dataset = NUM_CALIBRATION_SAMPLES // 2
+
+ds_ultrachat = load_dataset(
+    "HuggingFaceH4/ultrachat_200k",
+    split=f"train_sft[:{samples_per_dataset}]",
+)
+ds_nemotron = load_dataset(
+    "nvidia/Nemotron-Post-Training-Dataset-v2",
+    split=f"chat[:{samples_per_dataset}]",
+)
+
+# Both datasets share a "messages" column with the same chat format.
+# Keep only that column so we can concatenate them.
+ds_ultrachat = ds_ultrachat.select_columns(["messages"])
+ds_nemotron = ds_nemotron.select_columns(["messages"])
+ds = concatenate_datasets([ds_ultrachat, ds_nemotron])
+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 and scheme.
+# In this case, we:
+#   * quantize the weights to fp4 with per group 16 via ptq
+#   * calibrate a global_scale for activations, which will be used to
+#       quantize activations to fp4 on the fly
+recipe = QuantizationModifier(
+    targets="Linear",
+    scheme="NVFP4",
+    ignore=[
+        "lm_head",
+        "re:.*mlp.gate$",
+        "re:.*mlp.shared_expert_gate$",
+        "re:.*linear_attn.*",
+        "re:model\\.visual\\..*",
+    ],
+)
+
+# Apply quantization.
+# MoE calibration is now handled automatically by the pipeline.
+# We set `moe_calibrate_all_experts` to True to ensure all experts receive
+# calibration data. This temporarily updates the model definition to use
+# `CalibrationQwen3_5MoeSparseMoeBlock` (from `llmcompressor.modeling.qwen3_5_moe`)
+# which replaces the original `Qwen3_5MoeSparseMoeBlock` class.
+# This unfuses the 3D expert parameters into individual nn.Linear modules
+# so they can be targeted by quantization.
+# Feel free to update the definition under
+# llm-compressor/src/llmcompressor/modeling/qwen3_5_moe.py to play around with
+# this behavior and evaluate its impact on quantization performance.
+oneshot(
+    model=model,
+    dataset=ds,
+    recipe=recipe,
+    max_seq_length=MAX_SEQUENCE_LENGTH,
+    num_calibration_samples=NUM_CALIBRATION_SAMPLES,
+    moe_calibrate_all_experts=True,
+)
+
+
+print("\n\n")
+print("========== SAMPLE GENERATION ==============")
+try:
+    dispatch_for_generation(model)
+    input_ids = tokenizer("Hello my name is", return_tensors="pt").input_ids.to(
+        model.device
+    )
+    output = model.generate(input_ids, max_new_tokens=100)
+    print(tokenizer.decode(output[0]))
+except Exception as e:
+    print(f"Generation failed (non-fatal): {e}")
+print("==========================================\n\n")
+
+
+# Save to disk in compressed-tensors format.
+# MTP (multi-token prediction) weights are automatically preserved from
+# the source checkpoint during save, enabling speculative decoding with vLLM.
+SAVE_DIR = MODEL_ID.rstrip("/").split("/")[-1] + "-NVFP4"
+model.save_pretrained(SAVE_DIR, save_compressed=True)
+tokenizer.save_pretrained(SAVE_DIR)
+
+# Hot-fix: copy processor configs that save_pretrained doesn't bring over
+cache_dir = snapshot_download(MODEL_ID, allow_patterns=["*.json"])
+for filename in [
+    "preprocessor_config.json",
+    "video_preprocessor_config.json",
+]:
+    src = os.path.join(cache_dir, filename)
+    if os.path.exists(src):
+        shutil.copyfile(src, os.path.join(SAVE_DIR, filename))

src/llmcompressor/modeling/__init__.py

@@ -15,8 +15,9 @@
 from .glm_moe_dsa import CalibrationGlmMoeDsaMoE  # noqa: F401
 from .llama4 import SequentialLlama4TextMoe  # noqa: F401
 from .qwen3_moe import CalibrationQwen3MoeSparseMoeBlock  # noqa: F401
+from .qwen3_5_moe import CalibrationQwen3_5MoeSparseMoeBlock  # noqa: F401
 from .qwen3_vl_moe import CalibrateQwen3VLMoeTextSparseMoeBlock  # noqa: F401
 from .qwen3_next_moe import CalibrationQwen3NextSparseMoeBlock  # noqa: F401
-# TODO: add granite4, Qwen3Next
+# TODO: add granite4
 
 from .fuse import *

src/llmcompressor/modeling/qwen3_5_moe.py

@@ -0,0 +1,140 @@
+from __future__ import annotations
+
+from typing import TYPE_CHECKING
+
+import torch
+import torch.nn.functional as F
+
+from llmcompressor.modeling.moe_context import MoECalibrationModule
+from llmcompressor.utils.dev import skip_weights_initialize
+
+if TYPE_CHECKING:
+    from transformers.models.qwen3_5_moe.modeling_qwen3_5_moe import (
+        Qwen3_5MoeSparseMoeBlock,
+    )
+
+
+@MoECalibrationModule.register("Qwen3_5MoeSparseMoeBlock")
+class CalibrationQwen3_5MoeSparseMoeBlock(MoECalibrationModule):
+    """
+    Calibration version of Qwen3_5MoeSparseMoeBlock that unfuses 3D expert
+    parameters into individual MLP modules (nn.Linear) so they can be
+    individually quantized. Sends all tokens to all experts during calibration.
+
+    is_permanent = True because the unfused structure must persist for
+    quantization to target the individual nn.Linear expert weights.
+    """
+
+    is_permanent = True
+
+    def __init__(
+        self,
+        original: Qwen3_5MoeSparseMoeBlock,
+        config,
+        calibrate_all_experts: bool = True,
+    ):
+        super().__init__()
+        text_config = getattr(config, "text_config", config)
+
+        self.num_experts = text_config.num_experts
+        self.top_k = text_config.num_experts_per_tok
+        self.hidden_size = text_config.hidden_size
+
+        self.calibrate_all_experts = calibrate_all_experts
+        self.gate = original.gate
+        self.shared_expert = original.shared_expert
+        self.shared_expert_gate = original.shared_expert_gate
+        self.experts = SequentialQwen3_5MoeExperts(text_config, original.experts)
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        batch_size, sequence_length, hidden_dim = hidden_states.shape
+        hidden_states_reshaped = hidden_states.view(-1, hidden_dim)
+
+        # router: returns (router_logits, router_scores, router_indices)
+        _, routing_weights, selected_experts = self.gate(hidden_states_reshaped)
+
+        # expert mask: (num_experts, top_k, num_tokens)
+        expert_mask = F.one_hot(selected_experts, num_classes=self.num_experts).permute(
+            2, 1, 0
+        )
+
+        final_hidden_states = torch.zeros(
+            (batch_size * sequence_length, hidden_dim),
+            dtype=hidden_states.dtype,
+            device=hidden_states.device,
+        )
+
+        for expert_idx, expert_layer in enumerate(self.experts):
+            idx, token_idx = torch.where(expert_mask[expert_idx])
+
+            if self.calibrate_all_experts:
+                expert_out = expert_layer(hidden_states_reshaped)[token_idx]
+            else:
+                expert_out = expert_layer(hidden_states_reshaped[token_idx])
+
+            if len(token_idx) > 0:
+                current_hidden_states = (
+                    expert_out * routing_weights[token_idx, idx, None]
+                )
+                final_hidden_states.index_add_(
+                    0,
+                    token_idx,
+                    current_hidden_states.to(hidden_states.dtype),
+                )
+
+        # shared expert
+        shared_expert_output = self.shared_expert(hidden_states_reshaped)
+        shared_expert_output = (
+            F.sigmoid(self.shared_expert_gate(hidden_states_reshaped))
+            * shared_expert_output
+        )
+        final_hidden_states = final_hidden_states + shared_expert_output
+
+        final_hidden_states = final_hidden_states.reshape(
+            batch_size, sequence_length, hidden_dim
+        )
+        return final_hidden_states
+
+    def restore(self, original: torch.nn.Module) -> torch.nn.Module:
+        return self
+
+
+class SequentialQwen3_5MoeExperts(torch.nn.ModuleList):
+    """
+    Unfuses 3D expert parameter tensors into individual Qwen3_5MoeMLP modules
+    so that each expert's weights are nn.Linear and can be targeted by
+    quantization with targets="Linear".
+    """
+
+    def __init__(self, config, original):
+        from transformers.models.qwen3_5_moe.modeling_qwen3_5_moe import (
+            Qwen3_5MoeMLP,
+        )
+
+        self.num_experts = config.num_experts
+        intermediate_size = config.moe_intermediate_size
+
+        with skip_weights_initialize():
+            super().__init__(
+                [
+                    Qwen3_5MoeMLP(config, intermediate_size=intermediate_size)
+                    for _ in range(self.num_experts)
+                ]
+            )
+
+        gate_up_data = original.gate_up_proj.data  # [num_experts, 2*inter, hidden]
+        down_data = original.down_proj.data  # [num_experts, hidden, inter]
+
+        for i in range(self.num_experts):
+            gate_up = gate_up_data[i]  # [2*intermediate, hidden]
+            down = down_data[i]  # [hidden, intermediate]
+
+            # gate_up_proj stores [gate; up] stacked along dim 0
+            # nn.Linear weight is [out_features, in_features]
+            self[i].gate_proj.weight.data = (
+                gate_up[:intermediate_size, :].clone().contiguous()
+            )
+            self[i].up_proj.weight.data = (
+                gate_up[intermediate_size:, :].clone().contiguous()
+            )
+            self[i].down_proj.weight.data = down.clone().contiguous()

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

@@ -1,5 +1,6 @@
 import tqdm
 from compressed_tensors.utils import match_named_modules
+from loguru import logger
 
 from llmcompressor.core import Event, EventType, State
 from llmcompressor.modifiers import Modifier
@@ -111,3 +112,25 @@ def on_end(self, state: State, event: Event, **kwargs):
     def on_finalize(self, state: State, **kwargs) -> bool:
         if not self.ended_:
             self.on_end(state, None)
+
+        # Expand regex ignore patterns to explicit module names.
+        # QuantizationConfig.from_pretrained() builds the ignore list by
+        # checking module types, which misses non-standard modules matched
+        # by regex (e.g. MoE router modules that aren't nn.Linear).
+        # Store expanded names on the model so the save wrapper can ensure
+        # they appear in config.json.
+        regex_patterns = [p for p in self.ignore if p.startswith("re:")]
+        if regex_patterns:
+            expanded = {
+                name
+                for name, _ in match_named_modules(state.model, regex_patterns)
+            }
+            if expanded:
+                existing = getattr(
+                    state.model, "_quantization_expanded_ignore", set()
+                )
+                state.model._quantization_expanded_ignore = existing | expanded
+                logger.info(
+                    f"Expanded {len(regex_patterns)} regex ignore pattern(s) "
+                    f"to {len(expanded)} explicit module name(s)"
+                )