[NPU] Support DFlash speculative decoding for MiMo-V2.5-Pro (mxfp4) (#37565)

test/manual/ascend/llm_models/test_npu_mimo_v2_5_pro_fp4_dflash.py

@@ -0,0 +1,47 @@
+import unittest
+
+from sglang.test.ascend.gsm8k_ascend_mixin import GSM8KAscendMixin
+from sglang.test.ascend.test_ascend_utils import (
+    MIMO_V2_5_PRO_FP4_DFLASH_DRAFT_WEIGHTS_PATH,
+    MIMO_V2_5_PRO_FP4_DFLASH_WEIGHTS_PATH,
+)
+from sglang.test.test_utils import CustomTestCase
+
+
+class TestMiMoV25ProFP4GraphWithDFlash(GSM8KAscendMixin, CustomTestCase):
+    """Testcase: Verify the inference accuracy of MiMo-V2.5-Pro-FP4 on GSM8K with npu graph and DFlash speculative decoding.
+
+    [Test Category] Model
+    [Test Target] MiMo-V2.5-Pro-FP4-DFlash
+    [Test Config] Prefill+Decode, npu graph enabled, DFLASH speculative decoding, mxfp4 quantization, dp attention
+    """
+
+    model = MIMO_V2_5_PRO_FP4_DFLASH_WEIGHTS_PATH
+    accuracy = 0.9
+    other_args = [
+        "--trust-remote-code",
+        "--mem-fraction-static",
+        "0.87",
+        "--attention-backend",
+        "ascend",
+        "--tp-size",
+        "8",
+        "--moe-a2a-backend",
+        "deepep",
+        "--deepep-mode",
+        "auto",
+        "--speculative-algorithm",
+        "DFLASH",
+        "--speculative-draft-model-path",
+        MIMO_V2_5_PRO_FP4_DFLASH_DRAFT_WEIGHTS_PATH,
+        "--speculative-num-draft-tokens",
+        "8",
+        "--dp-size",
+        "2",
+        "--enable-dp-attention",
+        "--enable-dp-lm-head",
+    ]
+
+
+if __name__ == "__main__":
+    unittest.main()

test/registered/spec/dspark/test_dspark_stacked_ctx_kv_parity.py

@@ -58,6 +58,9 @@ def _make_attn(rope, *, eps=EPS, has_bias=False, quantized=False, g=None):
         k_norm.weight.copy_(torch.randn(HEAD_DIM, device=DEVICE, generator=g))
     attn.k_norm = k_norm
     attn.rotary_emb = rope
+    # Real DFlashAttention always sets v_scale in __init__; the mock must match
+    # so kv_proj_only's v_scale read does not AttributeError.
+    attn.v_scale = None
     for name in ("kv_proj_only", "apply_k_norm", "apply_k_rope"):
         setattr(attn, name, types.MethodType(getattr(DFlashAttention, name), attn))
     return attn