[Model][Dflash] Enable Dflash support for Qwen3NextForCausalLM targets (#45319)

Signed-off-by: Jonas I. Liechti <j-i-l@t4d.ch>

tests/models/registry.py

@@ -1425,6 +1425,14 @@ _SPECULATIVE_DECODING_EXAMPLE_MODELS = {
         max_model_len=8192,  # Reduce max len to ensure test runs in low-VRAM CI env
         max_num_seqs=32,
     ),
+    "DFlashQwen3NextDraftModel": _HfExamplesInfo(
+        "Qwen/Qwen3-Coder-Next",
+        speculative_model="z-lab/Qwen3-Coder-Next-DFlash",
+        use_original_num_layers=True,  # DFlash requires all layers
+        max_model_len=8192,  # Reduce for CI
+        max_num_seqs=32,
+        min_transformers_version="4.56.3",  # Required for Qwen3Next
+    ),
     # [Eagle]
     "EagleCohereForCausalLM": _HfExamplesInfo(
         "/host/engines/cohere-moe",

vllm/model_executor/models/qwen3_next.py

@@ -68,6 +68,7 @@ from .interfaces import (
     HasInnerState,
     IsHybrid,
     MixtureOfExperts,
+    SupportsEagle3,
     SupportsLoRA,
     SupportsPP,
 )
@@ -758,6 +759,7 @@ class Qwen3NextForCausalLM(
     SupportsPP,
     QwenNextMixtureOfExperts,
     IsHybrid,
+    SupportsEagle3,
 ):
     packed_modules_mapping = {
         "qkv_proj": [