Added sliding window attention support for qwen-eagle3 architecture (#47568)

Signed-off-by: shanjiaz <zsjwpianpian@gmail.com>
This commit is contained in:
shanjiaz
2026-07-13 20:20:44 +00:00
committed by GitHub
parent 7fc97042c3
commit 7dc447dda7
2 changed files with 15 additions and 0 deletions
@@ -43,11 +43,22 @@ class Qwen3Eagle3DecoderLayer(Qwen3DecoderLayer):
cache_config = vllm_config.cache_config
quant_config = get_draft_quant_config(vllm_config)
# Resolve per-layer sliding window from draft config
sliding_window = None
layer_types = getattr(config, "layer_types", None)
if (
layer_types
and layer_idx < len(layer_types)
and layer_types[layer_idx] == "sliding_attention"
):
sliding_window = getattr(config, "sliding_window", None)
super().__init__(
config=config,
cache_config=cache_config,
quant_config=quant_config,
prefix=prefix,
per_layer_sliding_window=sliding_window,
)
# First layer uses 2*hidden_size (embeds + hidden_states concatenated)