Skip to content

Official Inference Code of EAGLE-4

fix(vllm): stop requiring vLLM's varlen verify-graph flag

Seyfal Sultanovbf8df90246c5
diff --git a/eagle_harness/vllm/adapter.py b/eagle_harness/vllm/adapter.py
index 330b94a..b051e89 100644
--- a/eagle_harness/vllm/adapter.py
+++ b/eagle_harness/vllm/adapter.py
@@ -77,11 +77,6 @@ HOOK_MAP: Tuple[HookUse, ...] = (
" | ".join(f"{m}:LinearAttnVerifyExtension.forward_spec_rows" for m in VERIFY_EXTENSION_MODULES),
"gpu",
),
- HookUse(
- "H7 per-width verify graphs",
- "vllm.v1.worker.gpu.spec_decode.speculator:BaseSpeculator.uses_varlen_verify_graphs",
- "gpu",
- ),
HookUse(
"tree attention backend (H9 plan)", "vllm.v1.attention.backends.registry:AttentionBackendEnum.CUSTOM", "gpu"
),