| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | """Utility functions for Helion kernel management.""" |
| 4 | |
| 5 | import regex as re |
| 6 | import torch |
| 7 | |
| 8 | from vllm.logger import init_logger |
| 9 | from vllm.platforms import current_platform |
| 10 | |
| 11 | logger = init_logger(__name__) |
| 12 | |
| 13 | # Maps known variant GPU names (after lowercase/underscore normalization) |
| 14 | # to their canonical form. |
| 15 | # |
| 16 | # Names that are already canonical after normalization are NOT listed here. |
| 17 | # For example, "NVIDIA H200" normalizes to "nvidia_h200" which needs no |
| 18 | # further mapping, and AMD ROCm names like "AMD_Instinct_MI300X" come from |
| 19 | # a controlled lookup table in rocm.py and normalize cleanly to |
| 20 | # "amd_instinct_mi300x". Only names with variant suffixes (form factor, |
| 21 | # memory size, memory type, etc.) that should be stripped need entries. |
| 22 | # |
| 23 | # To add a new GPU variant: run `canonicalize_gpu_name()` without the alias |
| 24 | # to see the normalized name, then add a mapping here if it contains variant |
| 25 | # suffixes that should be stripped (e.g. Blackwell/Rubin variants). |
| 26 | _GPU_NAME_ALIASES: dict[str, str] = { |
| 27 | # H100 variants |
| 28 | "nvidia_h100_pcie": "nvidia_h100", |
| 29 | "nvidia_h100_sxm5": "nvidia_h100", |
| 30 | "nvidia_h100_80gb_hbm3": "nvidia_h100", |
| 31 | "nvidia_h100_nvl": "nvidia_h100", |
| 32 | # H200 variants |
| 33 | "nvidia_h200_nvl": "nvidia_h200", |
| 34 | "nvidia_h200_141gb_hbm3e": "nvidia_h200", |
| 35 | # A100 variants |
| 36 | "nvidia_a100_sxm4_80gb": "nvidia_a100", |
| 37 | "nvidia_a100_sxm4_40gb": "nvidia_a100", |
| 38 | "nvidia_a100_pcie_80gb": "nvidia_a100", |
| 39 | "nvidia_a100_pcie_40gb": "nvidia_a100", |
| 40 | "nvidia_a100_80gb_pcie": "nvidia_a100", |
| 41 | # V100 variants (Tesla-branded) |
| 42 | "tesla_v100_sxm2_32gb": "tesla_v100", |
| 43 | "tesla_v100_sxm2_16gb": "tesla_v100", |
| 44 | "tesla_v100_pcie_32gb": "tesla_v100", |
| 45 | "tesla_v100_pcie_16gb": "tesla_v100", |
| 46 | # AMD ROCm variants (from _ROCM_DEVICE_ID_NAME_MAP in rocm.py) |
| 47 | "amd_instinct_mi300x_hf": "amd_instinct_mi300x", |
| 48 | # ADD MORE HERE |
| 49 | } |
| 50 | |
| 51 | |
| 52 | def get_gpu_name(device_id: int | None = None) -> str: |
| 53 | if device_id is None: |
| 54 | logger.warning_once( |
| 55 | "get_gpu_name() called without device_id, defaulting to 0. " |
| 56 | "This may return the wrong device name in multi-node setups." |
| 57 | ) |
| 58 | device_id = 0 |
| 59 | return current_platform.get_device_name(device_id) |
| 60 | |
| 61 | |
| 62 | def canonicalize_gpu_name(name: str) -> str: |
| 63 | """Canonicalize GPU name for use as a platform identifier. |
| 64 | |
| 65 | Converts to lowercase, replaces separators with underscores, |
| 66 | and maps known variant names to their canonical form via _GPU_NAME_ALIASES. |
| 67 | e.g., "NVIDIA H100 80GB HBM3" -> "nvidia_h100" |
| 68 | "NVIDIA A100-SXM4-80GB" -> "nvidia_a100" |
| 69 | "AMD Instinct MI300X" -> "amd_instinct_mi300x" |
| 70 | """ |
| 71 | if not name or not name.strip(): |
| 72 | raise ValueError("GPU name cannot be empty") |
| 73 | name = re.sub(r"[\s/-]+", "_", name.lower()) |
| 74 | if name in _GPU_NAME_ALIASES: |
| 75 | return _GPU_NAME_ALIASES[name] |
| 76 | return name |
| 77 | |
| 78 | |
| 79 | def get_canonical_gpu_name(device_id: int | None = None) -> str: |
| 80 | return canonicalize_gpu_name(get_gpu_name(device_id)) |
| 81 | |
| 82 | |
| 83 | def get_fp8_dtype() -> torch.dtype: |
| 84 | return current_platform.fp8_dtype() |
| 85 | |
| 86 | |
| 87 | def get_int8_min_max() -> tuple[int, int]: |
| 88 | qtype_traits = torch.iinfo(torch.int8) |
| 89 | return qtype_traits.min, qtype_traits.max |
| 90 | |
| 91 | |
| 92 | def get_int8_min_scaling_factor() -> float: |
| 93 | return torch.finfo(torch.float32).eps |