From eab1cff5b0ca87ec415c3ca555f060933d0a6b72 Mon Sep 17 00:00:00 2001
From: yimdev <4mengy@gmail.com>
Date: Tue, 18 Aug 2026 16:43:12 +0800
Subject: [PATCH] Harden DeepSeek V3.2 fused kernel grids (#52381)

Signed-off-by: yimdev <5779256+yimdev@users.noreply.github.com>
Co-authored-by: mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
---
 .../test_fused_deepseek_v32_norm_rope.py      | 73 +++++++++++++++++++
 vllm/models/deepseek_v32/common/kernels.py    | 12 +--
 2 files changed, 79 insertions(+), 6 deletions(-)

diff --git a/tests/kernels/test_fused_deepseek_v32_norm_rope.py b/tests/kernels/test_fused_deepseek_v32_norm_rope.py
index 48a1f78edb..d72c1b31c4 100644
--- a/tests/kernels/test_fused_deepseek_v32_norm_rope.py
+++ b/tests/kernels/test_fused_deepseek_v32_norm_rope.py
@@ -362,6 +362,46 @@ def test_fused_norm_rope_ds_mla(num_tokens: int):
     assert (topk == 7).all(), "topk buffer should be untouched (no indexer)"
 
 
+def test_fused_norm_rope_supports_large_token_count():
+    """Keep the token count off CUDA grid-y at its 65,536-block boundary."""
+    num_tokens = 65536
+    dev = "cuda"
+    dtype = torch.bfloat16
+    positions = torch.zeros(num_tokens, device=dev, dtype=torch.int64)
+    q_c = torch.ones((num_tokens, 1), device=dev, dtype=dtype)
+    kv_c = torch.ones((num_tokens, 1), device=dev, dtype=dtype)
+    k_pe = torch.ones((num_tokens, 2), device=dev, dtype=dtype)
+    norm_w = torch.ones(1, device=dev, dtype=dtype)
+    cos_sin = torch.tensor([[1.0, 0.0]], device=dev, dtype=torch.float32)
+    topk = torch.empty((num_tokens, 1), device=dev, dtype=torch.int32)
+    slot_mapping = torch.arange(num_tokens, device=dev, dtype=torch.int64)
+    mla_cache = torch.empty((1, num_tokens, 3), device=dev, dtype=dtype)
+
+    q_out = K.fused_norm_rope(
+        positions,
+        q_c,
+        norm_w,
+        EPS,
+        kv_c,
+        norm_w,
+        EPS,
+        k_pe,
+        cos_sin,
+        None,
+        None,
+        None,
+        EPS,
+        None,
+        topk,
+        slot_mapping=slot_mapping,
+        mla_kv_cache=mla_cache,
+        has_indexer=False,
+    )
+
+    rows = torch.tensor([0, num_tokens - 1], device=dev)
+    assert_bf16(q_out[rows], rms_norm(q_c[rows], norm_w), "large-token q norm")
+
+
 # ── fused_q ──────────────────────────────────────────────────────────────────
 
 
@@ -539,6 +579,39 @@ def test_fused_q_bf16_query(num_tokens: int, has_indexer: bool):
         torch.testing.assert_close(iw_out, iw_ref, rtol=1e-3, atol=1e-3)
 
 
+def test_fused_q_triton_supports_large_token_count():
+    """Keep the token count off CUDA grid-y in the Triton fallback.
+
+    The minimal dimensions also bypass CuTeDSL on SM100.
+    """
+    num_tokens = 65536
+    dev = "cuda"
+    dtype = torch.bfloat16
+    positions = torch.zeros(num_tokens, device=dev, dtype=torch.int64)
+    q_pe = torch.ones((num_tokens, 1, 2), device=dev, dtype=dtype)
+    ql_nope = torch.ones((num_tokens, 1, 1), device=dev, dtype=dtype)
+    cos_sin = torch.tensor([[1.0, 0.0]], device=dev, dtype=torch.float32)
+    q_scale = torch.ones(1, device=dev, dtype=torch.float32)
+
+    _, _, mqa_q = K.fused_q(
+        positions,
+        q_pe,
+        cos_sin,
+        None,
+        None,
+        ql_nope,
+        q_scale,
+        None,
+        0.0,
+        0.0,
+        has_indexer=False,
+    )
+
+    rows = torch.tensor([0, num_tokens - 1], device=dev)
+    ref = torch.cat([ql_nope, q_pe], dim=-1).to(FP8)
+    assert_fp8(mqa_q[rows], ref[rows], "large-token fused Q")
+
+
 # ── fused_eh_norm (MTP) ──────────────────────────────────────────────────────
 
 
diff --git a/vllm/models/deepseek_v32/common/kernels.py b/vllm/models/deepseek_v32/common/kernels.py
index 2c6e895fdb..ec642409cb 100644
--- a/vllm/models/deepseek_v32/common/kernels.py
+++ b/vllm/models/deepseek_v32/common/kernels.py
@@ -152,8 +152,8 @@ def _fused_norm_rope_kernel(
     INDEX_ROPE_INTERLEAVE: tl.constexpr,
     USE_PDL: tl.constexpr,
 ):
-    pid = tl.program_id(0)
-    tok_idx = tl.program_id(1)
+    tok_idx = tl.program_id(0).to(tl.int64)
+    pid = tl.program_id(1)
     if USE_PDL:
         tl.extra.cuda.gdc_wait()
         tl.extra.cuda.gdc_launch_dependents()
@@ -479,7 +479,7 @@ def fused_norm_rope(
     if q_c_out is None:
         q_c_out = torch.empty_like(q_c)
     use_pdl = current_platform.is_arch_support_pdl()
-    _fused_norm_rope_kernel[(4, num_tokens)](
+    _fused_norm_rope_kernel[(num_tokens, 4)](
         positions,
         # Q RMS norm
         q_c,
@@ -594,8 +594,8 @@ def _fused_q_kernel(
     QUANTIZE_MQA: tl.constexpr,
     USE_PDL: tl.constexpr,
 ):
-    pid = tl.program_id(0)
-    tok_idx = tl.program_id(1)
+    tok_idx = tl.program_id(0).to(tl.int64)
+    pid = tl.program_id(1)
     head_idx = tl.program_id(2)
     if USE_PDL:
         tl.extra.cuda.gdc_wait()
@@ -869,7 +869,7 @@ def fused_q(
         return index_q_fp8, index_weights_out, mqa_q
 
     use_pdl = current_platform.is_arch_support_pdl()
-    _fused_q_kernel[(3, num_tokens, grid_heads)](
+    _fused_q_kernel[(num_tokens, 3, grid_heads)](
         positions,
         q_pe,
         q_pe.stride(0),