From a70d2d39842af7e6c8184e3e92e949472eb9e317 Mon Sep 17 00:00:00 2001 From: biondizzle Date: Sat, 16 May 2026 04:40:31 +0000 Subject: [PATCH] =?UTF-8?q?fix:=20clearer=20warmup=20message=20=E2=80=94?= =?UTF-8?q?=20'Compiling=20CuTeDSL=20NVFP4=20MegaMoE=20kernel'?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit --- vllm/patches/deepseek_v4.py | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/vllm/patches/deepseek_v4.py b/vllm/patches/deepseek_v4.py index 86de2081..2c1b5afc 100644 --- a/vllm/patches/deepseek_v4.py +++ b/vllm/patches/deepseek_v4.py @@ -508,13 +508,13 @@ class DeepseekV4MegaMoEExperts(nn.Module): # Warm up the CuTeDSL kernel (JIT compiles MLIR→PTX on first call). # This takes ~1-2 min but prevents the vLLM RPC timeout (5 min) from # killing the engine when the first inference request triggers compilation. - print(" Warming up CuTeDSL NVFP4 kernel (first call compiles)...", flush=True) + print(" Compiling CuTeDSL NVFP4 MegaMoE kernel (one-time JIT, ~1-2 min)...", flush=True) dummy_hidden = torch.randn(1, self.hidden_size, dtype=torch.bfloat16, device=self._cutedsl_runner.l1_fp4[0].device) dummy_ids = torch.zeros(1, 1, dtype=torch.int32, device=dummy_hidden.device) dummy_weights = torch.ones(1, 1, dtype=torch.float32, device=dummy_hidden.device) try: self._cutedsl_runner.run(dummy_hidden, dummy_weights, dummy_ids, expert_indices=[0]) - print(" CuTeDSL warmup complete ✓", flush=True) + print(" CuTeDSL NVFP4 MegaMoE kernel compiled ✓", flush=True) except Exception: print(" CuTeDSL warmup failed (will compile on first inference)", flush=True)