Skip to content

fix: exclude FP8 quantized weights from MXFP4 de-osc #1081

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
hann-wang wants to merge 2 commits into
base: main
Choose a base branch
from
Open

fix: exclude FP8 quantized weights from MXFP4 de-osc #1081

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
4be54cb
fix: exclude FP8 quantized weights from MXFP4 de-osc
hann-wang Sep 3, 2026
e5286cb
chore: lint
hann-wang Sep 3, 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
20 changes: 15 additions & 5 deletions primus/backends/megatron/core/extensions/primus_turbo.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1147,7 +1147,9 @@ def forward_internal(

if get_num_microbatches() == 1:
if is_first_microbatch:
self.quantized_weight_buffer = torch.empty(0, device=weight.device, dtype=torch.uint8)
self.quantized_weight_buffer = torch.empty(
0, device=weights.device, dtype=float4_e2m1fn_x2
)
Comment on lines +1150 to +1152
out = primus_turbo_torch.ops.gemm_fp4(
x,
weight,
Expand Down Expand Up @@ -1358,7 +1360,9 @@ def forward_internal(

if get_num_microbatches() == 1:
if is_first_microbatch:
self.quantized_weight_buffer = torch.empty(0, device=weight.device, dtype=torch.uint8)
self.quantized_weight_buffer = torch.empty(
0, device=weights.device, dtype=float4_e2m1fn_x2
)
out = primus_turbo_torch.ops.gemm_fp4(
x,
weight,
Expand Down Expand Up @@ -1562,7 +1566,9 @@ def forward_internal(

if get_num_microbatches() == 1:
if is_first_microbatch:
self.quantized_weight_buffer = torch.empty(0, device=weight.device, dtype=torch.uint8)
self.quantized_weight_buffer = torch.empty(
0, device=weights.device, dtype=float4_e2m1fn_x2
)
out = primus_turbo_torch.ops.gemm_fp4(
x,
weight,
Expand Down Expand Up @@ -1779,7 +1785,9 @@ def forward_internal(self, x, is_first_microbatch: bool = False):

if get_num_microbatches() == 1:
if is_first_microbatch:
self.quantized_weight_buffer = torch.empty(0, device=weight.device, dtype=torch.uint8)
self.quantized_weight_buffer = torch.empty(
0, device=weights.device, dtype=float4_e2m1fn_x2
)
out = primus_turbo_torch.ops.gemm_fp4(
inp,
weight,
Expand Down Expand Up @@ -2176,7 +2184,9 @@ def forward_internal(
# no persistent cache. Expose a lightweight runtime marker so
# weight de-oscillation can identify the grouped FP4 weight.
if is_first_microbatch:
self.quantized_weight_buffer = torch.empty(0, device=weights.device, dtype=torch.uint8)
self.quantized_weight_buffer = torch.empty(
0, device=weights.device, dtype=float4_e2m1fn_x2
)
out = primus_turbo_torch.ops.grouped_gemm_fp4(
x,
weights,
Expand Down
3 changes: 2 additions & 1 deletion primus/backends/megatron/core/optimizer/weight_deosc.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -351,7 +351,8 @@ def _build_eligible_ids(self, dist_opt) -> set:
for chunk in model_chunks:
modules = chunk.modules() if hasattr(chunk, "modules") else []
for module in modules:
if getattr(module, "quantized_weight_buffer", None) is None:
quantized_weight_buffer = getattr(module, "quantized_weight_buffer", None)
if quantized_weight_buffer is None or quantized_weight_buffer.dtype != _float4_e2m1fn_x2:
continue
Comment on lines +354 to 356
weight = getattr(module, "_parameters", {}).get("weight", None)
if weight is None:
Expand Down
Loading