Skip to content
Open
Show file tree
Hide file tree
Changes from all commits
Commits
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: 3 additions & 2 deletions src/maxtext/configs/types.py
Original file line number Diff line number Diff line change
Expand Up @@ -2137,8 +2137,9 @@ class Qwen3_8FlashNext(BaseModel):
qsa_attention_kernel: Literal["dense", "flash"] = Field(
"dense",
description=(
"Qwen3.8 QSA attention kernel: 'dense' materializes the score matrix, 'flash' dispatches through"
" Tokamax Splash Attention (requires max_target_length <= index_select_budget)."
"Qwen3.8 QSA attention kernel: 'dense' materializes the full score matrix; 'flash' dispatches through"
" Tokamax Splash Attention when s <= index_select_budget and tiled block-sparse FlashAttention with a"
" 32-bit packed block bitmap when s > index_select_budget."
),
)
use_ple_layer: bool = Field(False, description="Whether to use Per-Layer Embedding (PLE).")
Expand Down
4 changes: 2 additions & 2 deletions src/maxtext/m3/models/__init__.py
Original file line number Diff line number Diff line change
Expand Up @@ -77,8 +77,8 @@ def create_granitemoe_model(config, mesh=None, *, rngs, **kwargs):
"granitemoe": _FOUR_AXIS_MESH,
"qwen3-0.6b": modeling_qwen3.m3_sharding.DEFAULT_SHARDING_AXES,
"qwen3-vl-2b": _FOUR_AXIS_MESH,
"qwen3.8-flash-next": ("data", "fsdp", "fsdp_transpose", "tensor", "tensor_sequence", "expert"),
"qwen3.8-flash-next-mini": ("data", "fsdp", "fsdp_transpose", "tensor", "tensor_sequence", "expert"),
"qwen3.8-flash-next": modeling_qwen3.m3_sharding.DEFAULT_SHARDING_AXES,
"qwen3.8-flash-next-mini": modeling_qwen3.m3_sharding.DEFAULT_SHARDING_AXES,
}
MODEL_EXPLICIT_SHARDING_SUPPORT: frozenset[str] = frozenset({"qwen3-0.6b"})

Expand Down
Loading
Loading