minke.yu/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 12 Packages Projects Releases Wiki Activity

Fix DFlash sliding attention causality defaults (#34524)

Browse Source
This commit is contained in:
Mohammad Miadh Angkad
2026-08-11 23:05:55 -07:00
committed by GitHub
parent 687967c70d
commit 0dab252ffc
2 changed files with 14 additions and 11 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/srt/configs/muse_glimmer.py
+1
View File
@@ -28,6 +28,7 @@ _ARCH = "muse-glimmer"
class MuseGlimmerAssistantConfig(PretrainedConfig): class MuseGlimmerAssistantConfig(PretrainedConfig):
model_type = "muse_glimmer_assistant" model_type = "muse_glimmer_assistant"
is_causal = False
# The DFlash draft has no head; draft_worker_common borrows the target's. # The DFlash draft has no head; draft_worker_common borrows the target's.
vocab_size = None vocab_size = None
python/sglang/srt/models/dflash.py
+13 -11
View File
@@ -43,14 +43,13 @@ if _is_npu:
logger = logging.getLogger(__name__) logger = logging.getLogger(__name__)
def _get_dflash_attention_type(config) -> AttentionType: def _get_dflash_attention_type(config, *, default: AttentionType) -> AttentionType:
"""Bidirectional over the draft block unless the checkpoint says causal.""" """Honor explicit causality while preserving legacy layer defaults."""
text_config = getattr(config, "text_config", None) or config text_config = config.get_text_config()
return ( is_causal = getattr(text_config, "is_causal", None)
AttentionType.DECODER if is_causal is None:
if getattr(text_config, "is_causal", False) return default
else AttentionType.ENCODER_ONLY return AttentionType.DECODER if is_causal else AttentionType.ENCODER_ONLY
)
def _get_dflash_layer_attention_params( def _get_dflash_layer_attention_params(
@@ -67,12 +66,15 @@ def _get_dflash_layer_attention_params(
layer_type = layer_types[layer_id] layer_type = layer_types[layer_id]
if layer_type == "full_attention": if layer_type == "full_attention":
return -1, _get_dflash_attention_type(config) return -1, _get_dflash_attention_type(
config, default=AttentionType.ENCODER_ONLY
)
if layer_type == "sliding_attention": if layer_type == "sliding_attention":
# Windowing is orthogonal to causality (mask is p1 - p0 >= window).
sliding_window_size = get_dflash_attention_sliding_window_size(config) sliding_window_size = get_dflash_attention_sliding_window_size(config)
assert sliding_window_size is not None assert sliding_window_size is not None
return sliding_window_size, _get_dflash_attention_type(config) return sliding_window_size, _get_dflash_attention_type(
config, default=AttentionType.DECODER
)
raise ValueError( raise ValueError(
"Unsupported DFLASH draft layer type. " "Unsupported DFLASH draft layer type. "
f"layer_types[{layer_id}]={layer_type!r}." f"layer_types[{layer_id}]={layer_type!r}."