pip install vllm   # 0.20.1 — deep_gemm is not listed as a required dependency
vllm serve TinyLlama/TinyLlama-1.1B-Chat-v1.0 \
    --gpu-memory-utilization 0.85 --max-model-len 2048

(EngineCore pid=3838313) ERROR [core.py:1136] EngineCore failed to start.
  File ".../vllm/v1/engine/core.py", line 1110, in run_engine_core
      engine_core = EngineCoreProc(...)
  File ".../vllm/v1/engine/core.py", line 128, in __init__
      kv_cache_config = self._initialize_kv_caches(vllm_config)
  File ".../vllm/v1/worker/gpu_worker.py", line 586, in compile_or_warm_up_model
      kernel_warmup(self)
  File ".../vllm/model_executor/warmup/kernel_warmup.py", line 37, in kernel_warmup
      deep_gemm_warmup(model, max_tokens)
  File ".../vllm/model_executor/warmup/deep_gemm_warmup.py", line 364, in deep_gemm_warmup
      total = _count_warmup_iterations(model, max_tokens)
  File ".../vllm/model_executor/warmup/deep_gemm_warmup.py", line 342, in _count_warmup_iterations
      if _fp8_linear_may_use_deep_gemm(m):
  File ".../vllm/model_executor/warmup/deep_gemm_warmup.py", line 135, in _fp8_linear_may_use_deep_gemm
      block_size = get_mk_alignment_for_contiguous_layout()[0]
  File ".../vllm/utils/deep_gemm.py", line 266, in get_mk_alignment_for_contiguous_layout
      return _missing()
RuntimeError: DeepGEMM backend is not available or outdated.
Please install or update the `deep_gemm` to a newer version to enable FP8 kernels.

def _fp8_linear_may_use_deep_gemm(module: torch.nn.Module) -> bool:
    # FIXME: this logic is brittle and incorrect
    block_size = get_mk_alignment_for_contiguous_layout()[0]  # ← crash for ALL modules
    if not (
        isinstance(module, LinearBase)
        and isinstance(module.quant_method, Fp8LinearMethod)   # ← never reached
        ...
    ):
        return False

+from vllm.utils.deep_gemm import (
     fp8_gemm_nt,
     get_mk_alignment_for_contiguous_layout,
+    is_deep_gemm_supported,
     m_grouped_fp8_gemm_nt_contiguous,
 )

 def _fp8_linear_may_use_deep_gemm(module: torch.nn.Module) -> bool:
     """Return True if the module could be processed with DeepGEMM."""
+    if not is_deep_gemm_supported():
+        return False
     block_size = get_mk_alignment_for_contiguous_layout()[0]
     ...

INFO  [monitor.py] torch.compile took 10.57 s in total
INFO  [gpu_worker.py] Available KV cache memory: 13.54 GiB
INFO  [kv_cache_utils.py] GPU KV cache size: 645,472 tokens
# Engine started — no DeepGEMM crash
INFO  vLLM ready — TinyLlama/TinyLlama-1.1B-Chat-v1.0 on port 8001

nvidia-smi -L
GPU 0: NVIDIA H100 80GB HBM3 (UUID: GPU-7f57f032-34ab-a064-c37c-7a4f75601cc5)
  MIG 2g.20gb     Device  0: (UUID: MIG-377e0049-554c-540b-93c6-d0976f8426cb)
  MIG 1g.10gb     Device  1: (UUID: MIG-f39aa248-6f1b-5769-b119-8d650bb34b27)
GPU 1: NVIDIA H100 80GB HBM3 (UUID: GPU-8c7c9c30-f12f-20bf-e14e-d3c4c965af32)

# Exact command that triggered the crash
CUDA_VISIBLE_DEVICES="MIG-377e0049-554c-540b-93c6-d0976f8426cb" \
  vllm serve TinyLlama/TinyLlama-1.1B-Chat-v1.0 \
  --gpu-memory-utilization 0.85 --max-model-len 2048 --port 8001

  File ".../vllm/platforms/cuda.py", line 174, in has_device_capability
      return super().has_device_capability(capability, device_id)
  File ".../vllm/platforms/interface.py", line 141, in has_device_capability
      current_capability = cls.get_device_capability(device_id=device_id)
  File ".../vllm/platforms/cuda.py", line 158, in get_device_capability
      physical_device_id = device_id_to_physical_device_id(device_id)
  File ".../vllm/platforms/cuda.py", line 57, in device_id_to_physical_device_id
      return int(physical_device_id)
             ^^^^^^^^^^^^^^^^^^^^^^^
ValueError: invalid literal for int() with base 10: 'MIG-377e0049-554c-540b-93c6-d0976f8426cb'

RuntimeError: Engine process failed to start. See stack trace for the root cause.

# vllm/platforms/interface.py
device_ids = os.environ[cls.device_control_env_var].split(",")
physical_device_id = device_ids[device_id]
return int(physical_device_id)   # ← ValueError for "MIG-<uuid>" strings

import os, torch
os.environ["CUDA_VISIBLE_DEVICES"] = "MIG-377e0049-554c-540b-93c6-d0976f8426cb"
torch.cuda.device_count()      # → 1
torch.cuda.get_device_name(0)  # → "NVIDIA H100 80GB HBM3 MIG 2g.20gb"

-            return int(physical_device_id)
+            try:
+                return int(physical_device_id)
+            except ValueError:
+                # MIG device UUID (e.g. "MIG-<uuid>") — CUDA already remaps
+                # it to logical device 0, so device_id is the correct index.
+                return device_id

INFO  [cuda.py] Using FLASH_ATTN attention backend (FlashAttention v3)
INFO  [kv_cache_utils.py] GPU KV cache size: 645,472 tokens
INFO  vLLM ready — TinyLlama/TinyLlama-1.1B-Chat-v1.0 on port 8001