Skip to content

[megatron] fix: update patch for MLA flashattn forward #6005

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
wuxibin89 merged 1 commit into from
Apr 15, 2026
+17 −12
Merged

[megatron] fix: update patch for MLA flashattn forward #6005

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
29 changes: 17 additions & 12 deletions verl/models/mcore/patch.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -34,6 +34,7 @@ def apply_patch():
from packaging import version

mcore_ge_013 = version.parse(megatron.core.__version__) >= version.parse("0.13.0")
mcore_ge_0162 = version.parse(megatron.core.__version__) >= version.parse("0.16.2")

def patch_get_query_key_value_tensors(
self,
Expand Down Expand Up @@ -294,14 +295,17 @@ def patch_forward(
# core attention computation
# ==================================
# Need corresponding TE change
non_dsa_thd_qkv_format = (
packed_seq_params
and packed_seq_params.qkv_format == "thd"
orig_v_dim = value.shape[-1] if value is not None else None
thd_packed_seq = packed_seq_params is not None and packed_seq_params.qkv_format == "thd"
need_v_pad = (
thd_packed_seq
and getattr(self.config, "experimental_attention_variant", None) is None
and value is not None
and query.shape[-1] != orig_v_dim
)
v_dim = value.shape[-1]
if non_dsa_thd_qkv_format and query.shape[-1] != v_dim:
value = F.pad(value, [0, query.shape[-1] - v_dim])
if need_v_pad:
# Pad V so THD attention can run when Q/V head dims differ.
value = F.pad(value, [0, query.shape[-1] - orig_v_dim])
self.core_attention.hidden_size_per_attention_head_v = value.shape[-1]
if self.checkpoint_core_attention and self.training:
core_attn_out = self._checkpointed_attention_forward(
Expand All @@ -323,11 +327,11 @@ def patch_forward(
attn_mask_type=attn_mask_type,
**extra_kwargs,
)
if non_dsa_thd_qkv_format:
if core_attn_out.ndim == 2:
core_attn_out = core_attn_out.reshape(*core_attn_out.shape[:-1], -1, value.shape[-1])
if query.shape[-1] != v_dim:
core_attn_out = core_attn_out[..., :v_dim]
if thd_packed_seq:
if need_v_pad:
if core_attn_out.ndim == 2:
core_attn_out = core_attn_out.reshape(*core_attn_out.shape[:-1], -1, value.shape[-1])
core_attn_out = core_attn_out[..., :orig_v_dim]
# reshape to same output shape as unpacked case
# (t, np, hn) -> (t, b=1, h=np*hn)
# t is the pack size = sum (sq_i)
Expand All @@ -352,7 +356,8 @@ def patch_forward(
if not mcore_ge_013:
MLASelfAttention.get_query_key_value_tensors = patch_get_query_key_value_tensors

MultiLatentAttention.forward = patch_forward
if not mcore_ge_0162:
MultiLatentAttention.forward = patch_forward


def apply_patch_mbridge():
Expand Down
Loading