Skip to content

[Qwen3.5] Make linear attention INT8 quantization opt-in via extra_options #2094

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Closed
Jianhui Dai (daijh) wants to merge 2 commits into from
Closed

[Qwen3.5] Make linear attention INT8 quantization opt-in via extra_options #2094

Changes from 1 commit
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
e2fab0b
Make linear attention INT8 quantization opt-in via extra_options
daijh Apr 22, 2026
4bb4842
Update src/python/py/models/builders/qwen.py
daijh Apr 22, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
39 changes: 19 additions & 20 deletions src/python/py/models/builders/qwen.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1013,29 +1013,28 @@ def __init__(self, config, io_dtype, onnx_dtype, ep, cache_dir, extra_options):
# Disable fused RoPE in attention op - we apply mRoPE manually
self.attention_attrs["use_rope_in_attn"] = False

# Mixed-precision quantization for linear attention layers.
# Baseline: whole model INT4. Override linear attention layer nodes
# to INT8 for better accuracy with modest size increase.
# Optional mixed-precision quantization for linear attention layers.
# When enabled via extra_options linear_attention_int8=true, promotes
# linear attention layer nodes from INT4 to INT8 for better accuracy.
#
# Linear attention recurrence accumulates errors across the full sequence,
# unlike softmax attention which normalizes per-step.
int8_nodes = {}
for i, lt in enumerate(self.layer_types):
if lt == "linear_attention":
# All linear attention projections: INT8
for proj in ("in_proj_a", "in_proj_b", "in_proj_qkv", "in_proj_z", "out_proj"):
int8_nodes[f"/model/layers.{i}/linear_attn/{proj}/MatMul"] = {"bits": 8}
# MLP projections in linear attention layers: INT8
for proj in ("gate_proj", "up_proj", "down_proj"):
int8_nodes[f"/model/layers.{i}/mlp/{proj}/MatMul"] = {"bits": 8}

if int8_nodes:
algo_config = self.quant_attrs["int4"].get("algo_config")
if algo_config is not None and hasattr(algo_config, "customized_weight_config"):
algo_config.customized_weight_config.update(int8_nodes)
else:
algo_config = RTNWeightOnlyQuantConfig(customized_weight_config=int8_nodes)
self.quant_attrs["int4"]["algo_config"] = algo_config
if extra_options.get("linear_attention_int8", "false").lower() == "true":
Comment thread
daijh marked this conversation as resolved.
Outdated
int8_nodes = {}
for i, lt in enumerate(self.layer_types):
if lt == "linear_attention":
for proj in ("in_proj_a", "in_proj_b", "in_proj_qkv", "in_proj_z", "out_proj"):
int8_nodes[f"/model/layers.{i}/linear_attn/{proj}/MatMul"] = {"bits": 8}
for proj in ("gate_proj", "up_proj", "down_proj"):
int8_nodes[f"/model/layers.{i}/mlp/{proj}/MatMul"] = {"bits": 8}

if int8_nodes:
algo_config = self.quant_attrs["int4"].get("algo_config")
if algo_config is not None and hasattr(algo_config, "customized_weight_config"):
algo_config.customized_weight_config.update(int8_nodes)
else:
algo_config = RTNWeightOnlyQuantConfig(customized_weight_config=int8_nodes)
self.quant_attrs["int4"]["algo_config"] = algo_config

# Replace standard KV cache I/O with hybrid cache I/O
self._setup_hybrid_cache_io()
Expand Down
Loading