Skip to content

[Perf] [NPU] deepep: enable MXFP8 low-latency dispatch for Fp8MoEMethod FP4 experts #40519

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

Merged
sglang-npu-bot merged 6 commits into from
Oct 26, 2026
Merged

[Perf] [NPU] deepep: enable MXFP8 low-latency dispatch for Fp8MoEMethod FP4 experts #40519

Show file tree
Hide file tree
Changes from 2 commits
Commits
Show all changes
6 commits
Select commit Hold shift + click to select a range
12b643b
[NPU] Enable MXFP8 low-latency dispatch for Fp8MoEMethod FP4 experts
iridiumine Sep 23, 2026
4fe501e
Merge branch 'main' into pr-npu-fp4-mxfp8-dispatch
iridiumine Sep 28, 2026
8754895
Merge branch 'main' into pr-npu-fp4-mxfp8-dispatch
iridiumine Oct 8, 2026
5e4517c
Merge branch 'main' into pr-npu-fp4-mxfp8-dispatch
iridiumine Oct 12, 2026
2810e71
Merge branch 'main' into pr-npu-fp4-mxfp8-dispatch
iridiumine Oct 16, 2026
401e4a3
Merge branch 'main' into pr-npu-fp4-mxfp8-dispatch
iridiumine Oct 19, 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
5 changes: 1 addition & 4 deletions python/sglang/srt/hardware_backend/npu/quantization/fp4_moe_methods.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -171,10 +171,7 @@ def process_weights_after_loading(self, layer: torch.nn.Module) -> None:

if hasattr(layer, "dispatcher"):
layer.dispatcher.set_quant_config(
{
"normal_dispatcher_output_dtype": "bf16",
"low_latency_dispatcher_output_dtype": "mxfp8",
}
dict(self.w13_kernel.DISPATCHER_QUANT_CONFIG)
)
Comment on lines 172 to 175

@OrangeRedeng OrangeRedeng Oct 1, 2026

Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Hi! My PR include #26408 an alternative solution that makes it possible to automatically determine the type of output dispatcher (fixing a long‑broken strategy), and it is implemented for both normal and low‑latency modes.

@OrangeRedeng OrangeRedeng Oct 1, 2026

Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I’d also prefer a single interface for interacting with the dispatcher type, for example the change introduced in PR #39589 has now broken automatic detection of the dispatcher output dtype in the TP (tensor parallelism) case.


def apply(self, layer: torch.nn.Module, dispatch_output: "DispatchOutput"):
Expand Down
20 changes: 10 additions & 10 deletions python/sglang/srt/hardware_backend/npu/quantization/moe_methods.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -279,6 +279,15 @@ def _get_bias_args(
class NPUW4A8MXFP4MoEMethod(_NPUMoEMethodBase):
"""ModelSlim W4A8 MoE with packed MXFP4 weights and MXFP8 activations."""

# Dispatch wire format required by the W4A8 MXFP GMMs: low-latency
# dispatch quantizes activations to MXFP8 (E8M0 block scales); normal
# dispatch stays BF16 because A5 MXFP8 normal dispatch is intranode-only,
# which also preserves multi-node prefill when ``deepep-mode=auto``.
DISPATCHER_QUANT_CONFIG = {
"normal_dispatcher_output_dtype": "bf16",
"low_latency_dispatcher_output_dtype": "mxfp8",
}

def __init__(self, dynamic_quant_kwargs=_DEFAULT_DYNAMIC_QUANT):
super().__init__(quant_config=None)
self.dynamic_quant_kwargs = dynamic_quant_kwargs
Expand Down Expand Up @@ -308,17 +317,8 @@ def process_weights_after_loading(
weight.data, weight_scale.data
)

# A5 DeepEP low-latency dispatch quantizes the valid received rows to
# MXFP8 and returns the matching E8M0 block scales. Keep normal mode
# in BF16 because A5 MXFP8 normal dispatch is intranode-only; this also
# preserves multi-node prefill when ``deepep-mode=auto``.
if weight_prefix == "w13" and hasattr(layer, "dispatcher"):
layer.dispatcher.set_quant_config(
{
"normal_dispatcher_output_dtype": "bf16",
"low_latency_dispatcher_output_dtype": "mxfp8",
}
)
layer.dispatcher.set_quant_config(dict(self.DISPATCHER_QUANT_CONFIG))

def apply(
self,
Expand Down
11 changes: 10 additions & 1 deletion python/sglang/srt/layers/quantization/fp8.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2561,8 +2561,17 @@ def process_weights_after_loading(self, layer: Module) -> None:
self._prepare_hpc_ops_weights(layer)

if hasattr(layer, "dispatcher"):
# Execution kernels may declare the dispatch wire format they
# need (DISPATCHER_QUANT_CONFIG, e.g. the NPU W4A8 MXFP GMMs
# consume MXFP8 activations); otherwise report the checkpoint
# dtype and let the dispatcher pick.
kernel_quant_config = getattr(
getattr(layer, "w13_kernel", None), "DISPATCHER_QUANT_CONFIG", None
)
layer.dispatcher.set_quant_config(
{
kernel_quant_config
if kernel_quant_config is not None
else {
"weight_dtype": layer.w13_weight.dtype,
"use_mxfp8": self.use_mxfp8,
}
Expand Down
Loading