Skip to content

Add FSDP v2 MXFP8 mixed precision support #6

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
shjwudp merged 4 commits into from
May 14, 2026
Merged

Add FSDP v2 MXFP8 mixed precision support #6

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
8a2515b
Add FSDP v2 MXFP8 mixed precision support
Autumn1998 May 12, 2026
3a80853
Align FSDP v2 MXFP8 weight handling
Autumn1998 May 14, 2026
876a87b
Merge mfsdp refactor main into FSDP v2 MXFP8
Autumn1998 May 14, 2026
4e219fd
Clean up FSDP v2 bucket allocator ownership
Autumn1998 May 14, 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
7 changes: 7 additions & 0 deletions megatron/core/distributed/fsdp/mcore_fsdp_adapter.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -53,6 +53,7 @@
)
from megatron.core.distributed.fsdp.src.megatron_fsdp.v2 import FSDPModule
from megatron.core.distributed.fsdp.src.megatron_fsdp.v2.mixed_precision import (
FullyShardFP8Policy,
FullyShardMixedPrecisionPolicy,
)

Expand Down Expand Up @@ -251,6 +252,12 @@ def _init_with_fully_shard(
if ddp_config.grad_reduce_in_fp32
else ddp_config.megatron_fsdp_grad_comm_dtype
),
use_decoupled_grad=ddp_config.megatron_fsdp_use_decoupled_grad,
fp8=FullyShardFP8Policy(
enabled=ddp_config.fp8_param_gather,
recipe=config.fp8_recipe,
keep_transpose_cache=ddp_config.keep_fp8_transpose_cache,
),
)
kwargs = {
"mp_policy": fully_shard_mp_policy,
Expand Down
8 changes: 5 additions & 3 deletions megatron/core/distributed/fsdp/src/megatron_fsdp/v2/README.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -8,10 +8,12 @@ This directory contains the **fully_shard_v2** implementation — a PyTorch FSDP
v2/
├── README.md # This file
├── __init__.py # Public exports
├── fully_shard.py # Core fully_shard() API + FSDPModule class
├── fully_shard.py # Public fully_shard() API
├── fsdp_module.py # FSDPModule runtime state and methods
├── hooks.py # Forward/backward hook registration
├── param_group.py # ParameterGroup — groups params with shared buffers
├── dp_buffer.py # DataParallelBuffer — flat buffer management
├── allocator.py # TemporaryBucketAllocator — temp buffer reuse
├── allocator.py # BucketAllocator implementations
├── utils.py # Internal utility functions
└── design.md # Detailed design documentation (overlap, memory, sync)
```
Expand Down Expand Up @@ -123,4 +125,4 @@ torchrun --nproc_per_node=2 examples/megatron_fsdp/fsdp_toy.py \
```bash
# Run FSDP2 API tests
pytest -xvs tests/unit_tests/distributed/megatron_fsdp/test_mcore_fsdp_fully_shard_v2_api.py
```
```
Loading