Skip to content

[mxfp8 training] add new configurable params now exposed by torchao #2251

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
danielvegamyhre wants to merge 1 commit into
base: danielvegamyhre/stack/4
Choose a base branch
from
Open

[mxfp8 training] add new configurable params now exposed by torchao #2251

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
6 changes: 6 additions & 0 deletions torchtitan/components/quantization/mx.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -53,12 +53,18 @@ def __init__(self, job_config: JobConfig, parallel_dims: ParallelDims):

# Configure MXFP8
from torchao.prototype.mx_formats.config import (
MXFP8Dim0CastKernelChoice,
MXFP8Dim1CastKernelChoice,
MXLinearConfig as TorchAOMXLinearConfig,
)

mx_job_config = job_config.quantize.linear.mx
config = TorchAOMXLinearConfig.from_recipe_name(mx_job_config.recipe_name)
# pyrefly: ignore [missing-attribute]
config.mxfp8_dim0_cast_kernel_choice = MXFP8Dim0CastKernelChoice[
mx_job_config.mxfp8_dim0_cast_kernel_choice.upper()
]
# pyrefly: ignore [missing-attribute]
config.mxfp8_dim1_cast_kernel_choice = MXFP8Dim1CastKernelChoice[
mx_job_config.mxfp8_dim1_cast_kernel_choice.upper()
]
Expand Down
12 changes: 11 additions & 1 deletion torchtitan/config/job_config.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -757,7 +757,17 @@ class Float8GroupedMM:

@dataclass
class MXLinear:
mxfp8_dim1_cast_kernel_choice: Literal["triton", "cuda", "torch"] = "triton"
mxfp8_dim0_cast_kernel_choice: Literal["triton", "torch"] = "triton"
Copy link
Contributor

@tianyu-l tianyu-l Jan 24, 2026

Choose a reason for hiding this comment

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

default recipe is mxfp8_cublas for which torch kernel is recommended -- can we make the default consistent?

"""
Temp work around for inductor performance gap.

* triton is recommended for best performance for recipe "mxfp8_cublas_rceil" (rceil scale rounding mode)
* torch is recommended for best performance for recipe "mxfp8_cublas" (floor scale rounding mode)

Example: --quantize.linear.mx.mxfp8_dim0_cast_kernel_choice="torch"
"""

mxfp8_dim1_cast_kernel_choice: Literal["triton", "cuda", "torch"] = "cuda"
Copy link
Contributor

@tianyu-l tianyu-l Jan 24, 2026

Choose a reason for hiding this comment

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

do we need to keep triton and torch? If so, when are they recommended?

"""
Temp work around for inductor performance gap.

Expand Down
4 changes: 0 additions & 4 deletions torchtitan/models/llama4/infra/parallelize.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -639,10 +639,6 @@ def apply_compile(model: nn.Module, compile_config: CompileConfig, ep_enabled: b
# by wrapping each submod's forward instead of their __call__
moe = submod
for attr_name, submod in moe.named_children():
if attr_name == "experts":
Copy link
Contributor

@tianyu-l tianyu-l Jan 24, 2026

Choose a reason for hiding this comment

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

is this issue gone?

# NOTE: We don't compile token dispatch and token combine due to an issue on B200:
# https://github.com/pytorch/torchtitan/issues/1940
continue
setattr(
moe,
attr_name,
Expand Down
Loading