Skip to content

Qwen3-MoE: fused-expert quantisation, BF16 router and expert-wise GPTQ #324

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

Open
Shreyas8612 wants to merge 4 commits into
base: feat/phase-split-quantisation
Choose a base branch
from
Open

Qwen3-MoE: fused-expert quantisation, BF16 router and expert-wise GPTQ #324

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
30e8182
Add fused Qwen3-MoE experts, BF16 router and phase-aware attention
Sep 16, 2026
3fb8809
Wire Qwen3-MoE modules into replacement, phase hooks and rotation
Sep 16, 2026
ec32347
Extend GPTQ to fused Qwen3-MoE experts and sharded device maps
Sep 16, 2026
2b19da7
Add CPU tests for Qwen3-MoE phase quantisation and the router lane
Sep 16, 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
24 changes: 24 additions & 0 deletions src/chop/nn/quantized/modules/__init__.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -40,10 +40,22 @@

from .qwen3_moe import (
Qwen3MoeAttentionMXFP,
Qwen3MoeAttentionMXFPRotate,
Qwen3MoeAttentionMXInt,
Qwen3MoeAttentionMXIntRotate,
Qwen3MoeMLPMXFP,
Qwen3MoeMLPMXInt,
Qwen3MoeExpertsMXFP,
Qwen3MoeExpertsMXInt,
Qwen3MoeSparseMoeBlockBF16Router,
Qwen3MoeSparseMoeBlockMinifloat,
Qwen3MoeTopKRouterBF16,
Qwen3MoeTopKRouterMX,
Qwen3MoeTopKRouterMXInt8,
Qwen3MoeTopKRouterE4M3,
Qwen3MoeTopKRouterE5M2,
Qwen3MoeRMSNormMinifloat,
Qwen3MoeDecoderLayerMinifloat,
)

from .glm4_moe import (
Expand Down Expand Up @@ -399,11 +411,23 @@
}

quantized_qwen3_moe_module_map = {
"qwen3_moe_decoder_layer_minifloat": Qwen3MoeDecoderLayerMinifloat,
"qwen3_moe_self_attention_mxfp": Qwen3MoeAttentionMXFP,
"qwen3_moe_self_attention_mxfp_rotate": Qwen3MoeAttentionMXFPRotate,
"qwen3_moe_self_attention_mxint": Qwen3MoeAttentionMXInt,
"qwen3_moe_self_attention_mxint_rotate": Qwen3MoeAttentionMXIntRotate,
"qwen3_moe_mlp_mxfp": Qwen3MoeMLPMXFP,
"qwen3_moe_mlp_mxint": Qwen3MoeMLPMXInt,
"qwen3_moe_experts_mxfp": Qwen3MoeExpertsMXFP,
"qwen3_moe_experts_mxint": Qwen3MoeExpertsMXInt,
"qwen3_moe_sparse_block_bf16_router": Qwen3MoeSparseMoeBlockBF16Router,
"qwen3_moe_sparse_block_minifloat": Qwen3MoeSparseMoeBlockMinifloat,
"qwen3_moe_router_bf16": Qwen3MoeTopKRouterBF16,
"qwen3_moe_router_mx": Qwen3MoeTopKRouterMX,
"qwen3_moe_router_mxint8": Qwen3MoeTopKRouterMXInt8,
"qwen3_moe_router_e4m3": Qwen3MoeTopKRouterE4M3,
"qwen3_moe_router_e5m2": Qwen3MoeTopKRouterE5M2,
"qwen3_moe_rms_norm_minifloat": Qwen3MoeRMSNormMinifloat,
}

quantized_glm4_moe_module_map = {
Expand Down
8 changes: 8 additions & 0 deletions src/chop/nn/quantized/modules/phase_config.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -57,6 +57,14 @@
DECODE_FP_WEIGHT_ATTR = "_mase_decode_weight_fp"
DECODE_FP_BIAS_ATTR = "_mase_decode_bias_fp"

# Qwen3-MoE stores all routed experts in two fused Parameters rather than
# child Linear modules. These attributes mirror the linear handoff contract
# for the fused gate/up and down tensors.
GPTQ_DECODE_EXPERT_GATE_UP_ATTR = "_mase_gptq_gate_up_proj_decode"
GPTQ_DECODE_EXPERT_DOWN_ATTR = "_mase_gptq_down_proj_decode"
DECODE_FP_EXPERT_GATE_UP_ATTR = "_mase_decode_gate_up_proj_fp"
DECODE_FP_EXPERT_DOWN_ATTR = "_mase_decode_down_proj_fp"

_VALID_DECODE_POLICIES = ("fp_only", "quantized")
_VALID_KV_HANDOFFS = ("decode_format", "fp")

Expand Down
20 changes: 20 additions & 0 deletions src/chop/nn/quantized/modules/qwen3_moe/__init__.py
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,6 +1,26 @@
from .compat import SUPPORTED_TRANSFORMERS_VERSION, require_qwen3_moe_fused_abi
from .attention import (
Qwen3MoeAttentionMXFP,
Qwen3MoeAttentionMXFPRotate,
Qwen3MoeAttentionMXInt,
Qwen3MoeAttentionMXIntRotate,
)
from .mlp import Qwen3MoeMLPMXFP, Qwen3MoeMLPMXInt
from .experts import Qwen3MoeExpertsMXFP, Qwen3MoeExpertsMXInt
from .router import (
Qwen3MoeSparseMoeBlockBF16Router,
Qwen3MoeSparseMoeBlockMinifloat,
Qwen3MoeTopKRouterBF16,
)
from .router_precision import (
Qwen3MoeTopKRouterE4M3,
Qwen3MoeTopKRouterE5M2,
Qwen3MoeTopKRouterMX,
Qwen3MoeTopKRouterMXInt8,
ROUTER_MX_FORMATS,
ROUTER_PRECISION_SCHEMA,
router_decode_config,
router_phase_config,
)
from .rms_norm import Qwen3MoeRMSNormMinifloat
from .decoder_layer import Qwen3MoeDecoderLayerMinifloat
Loading