adapt-new-llm
GitHub指导适配 AutoRound 以支持新 LLM 架构,解决量化失败、块检测异常、MoE 模型处理及自定义前向传播等问题。
Trigger Scenarios
Install
npx skills add intel/auto-round --skill adapt-new-llm -g -y
SKILL.md
Frontmatter
{
"name": "adapt-new-llm",
"description": "Adapt AutoRound to support a new LLM architecture that doesn't work out-of-the-box. Use when quantization fails for a new model type, block detection doesn't find layers, MoE models need unfusing, custom forward passes are needed, or non-standard linear layer types need handling."
}
Adapting AutoRound for a New LLM Architecture
Overview
Most standard Transformers-based LLMs work with AutoRound out-of-the-box. This skill covers what to do when a new model architecture requires code changes. The need for adaptation typically arises from:
- Non-standard layer hierarchy (block detection fails)
- Fused Mixture-of-Experts (MoE) weights
- Non-standard linear layer types (not
nn.LinearorConv1D) - Complex multi-component architectures (multimodal routing)
- Shared cache keys or position embeddings
Step 0: Diagnose the Problem
Try quantizing the model first:
from auto_round import AutoRound
ar = AutoRound("your-org/your-model", scheme="W4A16", iters=2, nsamples=2)
ar.quantize_and_save(output_dir="./test_output", format="auto_round")
Common failure modes and their fixes:
| Error / Symptom | Root Cause | Fix Section |
|---|---|---|
| "No quantizable layers found" | Block detection failed | Step 1 |
| "Quantized 0/N layers" | Layers not nn.Linear/Conv1D |
Step 4 |
| Shape mismatch in MoE layers | Fused expert weights | Step 2 |
| Wrong outputs / calibration diverges | Forward pass not exercised correctly | Step 3 |
| Cache key errors (Gemma3-style) | Shared position embeddings | Step 5 |
Step 1: Fix Block Detection
AutoRound discovers quantizable blocks via get_block_names() which searches
recursively for nn.ModuleList instances. If your model has a non-standard
layer hierarchy, block detection may fail.
Check current detection
from auto_round.utils import get_block_names
model = ... # loaded model
print(get_block_names(model))
Option A: Use to_quant_block_names parameter
For simple cases, override block names without code changes:
ar = AutoRound(
model,
to_quant_block_names="model.decoder.layers", # explicit path
)
Option B: Register in SPECIAL_MULTIMODAL_BLOCK
For multimodal or multi-component models, add a custom block handler in
auto_round/special_model_handler.py:
def _get_your_model_multimodal_block(model, quant_vision=False):
"""Get block names for YourModel.
YourModel structure:
- encoder.layers: encoder blocks
- decoder.layers: decoder blocks
"""
block_names = []
if quant_vision and hasattr(model, "encoder"):
block_names.append([f"encoder.layers.{i}" for i in range(len(model.encoder.layers))])
block_names.append([f"decoder.layers.{i}" for i in range(len(model.decoder.layers))])
return block_names
# Register: key must match model.config.model_type
SPECIAL_MULTIMODAL_BLOCK["your_model_type"] = _get_your_model_multimodal_block
Also add to support lists if applicable:
# If text-only calibration works for this multimodal model:
SUPPORT_ONLY_TEXT_MODELS.append("your_model_type")
# If batch_size must be limited:
mllms_with_limited_bs = (..., "your_model_type")
Step 2: Handle MoE (Mixture-of-Experts) Models
MoE models often have fused 3D expert weights (shape
[num_experts, hidden, intermediate]) that must be "unfused" into per-expert
nn.Linear layers for quantization.
Check if auto-handled
Transformers >= 5.0 has a linear_loop experts interface that auto-unfuses
most MoE models. Test first — it may just work.
Register custom unfusing
If auto-unfusing fails, create a custom module in
auto_round/modeling/fused_moe/:
1. Create auto_round/modeling/fused_moe/your_moe.py:
"""Unfuse fused MoE weights for YourModel."""
import torch
import torch.nn as nn
from auto_round.modeling.fused_moe.replace_modules import register_replacement
@register_replacement("YourMoELayer")
def replace_your_moe_layer(module, name, model):
"""Replace FusedMoE with per-expert nn.Linear layers."""
experts = nn.ModuleList()
for i in range(module.num_experts):
linear = nn.Linear(module.hidden_size, module.intermediate_size, bias=False)
linear.weight.data = module.weight[i].clone()
experts.append(linear)
return experts
2. Register in BUILTIN_MODULES:
Edit auto_round/modeling/fused_moe/replace_modules.py:
BUILTIN_MODULES["your_model_type"] = LazyImport("auto_round.modeling.fused_moe.your_moe")
Existing MoE implementations
| Model Type | File | Pattern |
|---|---|---|
llama4 |
fused_moe/llama4.py |
Custom replacement for no use_experts_implementation |
deepseek_v2 |
fused_moe/deepseek_v2.py |
q_scale calibration for Gaudi |
step3p5 |
fused_moe/step3_5_moe.py |
Splits fused MoELinear |
qwen3_omni_moe |
fused_moe/qwen3_omni.py |
Thinker + talker MoE |
Step 3: Add Custom Forward Pass
Some models have non-standard forward passes that don't get calibrated correctly
with the default model.forward(). This is common for multi-component
architectures.
Edit _handle_special_model() in auto_round/special_model_handler.py:
def _your_model_forward(model, **kwargs):
"""Custom forward that routes through all quantizable components."""
# Example: route through both encoder and decoder
encoder_output = model.encoder(**kwargs)
decoder_output = model.decoder(encoder_output, **kwargs)
return decoder_output
def _handle_special_model(model):
...
if hasattr(model, "config") and model.config.model_type == "your_model_type":
from functools import partial
model.forward = partial(_your_model_forward, model)
return model
When is this needed?
- Model has multiple sub-models (thinker/talker, encoder/decoder)
- Default forward doesn't exercise all quantizable layers
- Model needs special input preprocessing during calibration
Existing examples
| Model | Custom Forward | Purpose |
|---|---|---|
deepseek_vl_v2 |
_deepseek_vl2_forward |
Route through language component |
qwen2_5_omni |
_qwen2_5_omni_forward |
Route through thinker → talker |
qwen3_omni_moe |
_qwen3_omni_moe_forward |
Handle MoE routing in omni model |
Step 4: Handle Non-Standard Linear Layers
AutoRound quantizes these layer types by default:
# auto_round/utils/common.py
SUPPORTED_LAYER_TYPES = (torch.nn.Linear, transformers.pytorch_utils.Conv1D)
INNER_SUPPORTED_LAYER_TYPES = ("FP8Linear",) # matched by class name string
If your model uses a custom linear type (e.g., QuantizedLinear, FP8Linear),
it won't be quantized unless registered.
Option A: String-based matching
INNER_SUPPORTED_LAYER_TYPES matches by class name string — useful for
external classes that can't be imported directly:
INNER_SUPPORTED_LAYER_TYPES = ("FP8Linear", "YourCustomLinear")
Option B: Type-based registration
If you can import the class:
from your_library import YourLinear
SUPPORTED_LAYER_TYPES = SUPPORTED_LAYER_TYPES + (YourLinear,)
Step 5: Handle Shared Cache Keys
Some models share tensors across blocks during inference (e.g., Gemma3's rotary position embeddings). These must be declared so the calibration cache doesn't duplicate or corrupt them.
Edit SPECIAL_SHARED_CACHE_KEYS in auto_round/special_model_handler.py:
SPECIAL_SHARED_CACHE_KEYS["YourModelForCausalLM"] = ("shared_position_embeddings", "shared_rope")
The key is the class name of the model (not model_type).
Step 6: Test
def test_your_model_quantization():
ar = AutoRound(
"your-org/your-model",
scheme="W4A16",
iters=2,
nsamples=2,
batch_size=2,
)
compressed_model, layer_config = ar.quantize()
# Verify layers were quantized
assert len(layer_config) > 0, "No layers were quantized"
ar.save_quantized(output_dir="./tmp_your_model", format="auto_round")
# Verify inference works
from auto_round.utils import model_infer
output = model_infer(compressed_model, tokenizer, "Hello world")
assert output is not None
Step 7: Update Documentation
- Add model to supported list in
README.md - Update
README_CN.mdwith equivalent Chinese content - Add example script if the model has notable differences
Checklist
-
get_block_names()finds all quantizable blocks - MoE layers (if any) are unfused correctly
-
calib()runs without shape errors - All target layers are quantized (check "Quantized X/Y layers" log)
- Forward pass exercises all quantizable components
- Quantized model produces valid outputs
- Export to target format works
- README.md + README_CN.md updated
Key Files
| File | Purpose |
|---|---|
auto_round/special_model_handler.py |
Block handlers, custom forwards, shared cache keys |
auto_round/modeling/fused_moe/replace_modules.py |
MoE unfusing registry (BUILTIN_MODULES) |
auto_round/utils/common.py |
SUPPORTED_LAYER_TYPES, INNER_SUPPORTED_LAYER_TYPES |
auto_round/utils/model.py |
get_block_names(), is_mllm_model(), model loading |
auto_round/compressors/data_driven.py |
New-architecture quantization loop and block scheduling |
auto_round/algorithms/quantization/base.py |
Quantizer block execution, sampling, and diffusion output configs |
auto_round/calibration/llm.py |
LLM calibration data collection and calib() flow |
auto_round/autoround.py |
AutoRound factory — model type routing logic |
Version History
- 8884cac Current 2026-08-20 04:27
- 3cae4b6 2026-07-25 10:12


