| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | import importlib.metadata |
| 5 | import os |
| 6 | import platform |
| 7 | from datetime import timedelta |
| 8 | from functools import cache, lru_cache, wraps |
| 9 | from typing import TYPE_CHECKING |
| 10 | |
| 11 | import regex as re |
| 12 | import torch |
| 13 | from torch.distributed import PrefixStore, ProcessGroup |
| 14 | from torch.distributed.distributed_c10d import is_nccl_available |
| 15 | |
| 16 | import vllm.envs as envs |
| 17 | from vllm.logger import init_logger |
| 18 | from vllm.v1.attention.backends.registry import AttentionBackendEnum |
| 19 | |
| 20 | from .interface import DeviceCapability, Platform, PlatformEnum, in_wsl |
| 21 | |
| 22 | if TYPE_CHECKING: |
| 23 | from vllm.config import VllmConfig |
| 24 | from vllm.config.kernel import IrOpPriorityConfig |
| 25 | from vllm.utils.argparse_utils import FlexibleArgumentParser |
| 26 | from vllm.v1.attention.backend import AttentionBackend |
| 27 | from vllm.v1.attention.selector import AttentionSelectorConfig |
| 28 | |
| 29 | logger = init_logger(__name__) |
| 30 | |
| 31 | _KV_CACHE_DTYPE_REASON = "kv_cache_dtype not supported" |
| 32 | _TURBOQUANT_LAYOUT_REASON = "no KV cache layout in common with TURBOQUANT" |
| 33 | |
| 34 | try: |
| 35 | from amdsmi import ( |
| 36 | AmdSmiException, |
| 37 | AmdSmiMemoryType, |
| 38 | amdsmi_get_gpu_asic_info, |
| 39 | amdsmi_get_gpu_device_uuid, |
| 40 | amdsmi_get_gpu_memory_total, |
| 41 | amdsmi_get_processor_handles, |
| 42 | amdsmi_init, |
| 43 | amdsmi_shut_down, |
| 44 | amdsmi_topo_get_link_type, |
| 45 | amdsmi_topo_get_numa_node_number, |
| 46 | ) |
| 47 | except ImportError as e: |
| 48 | logger.warning("Failed to import from amdsmi with %r", e) |
| 49 | |
| 50 | try: |
| 51 | import vllm._C # noqa: F401 |
| 52 | except ImportError as e: |
| 53 | logger.warning("Failed to import from vllm._C with %r", e) |
| 54 | |
| 55 | # import custom ops, trigger op registration |
| 56 | try: |
| 57 | import vllm._C_stable_libtorch # noqa: F401 |
| 58 | except ImportError as e: |
| 59 | logger.warning("Failed to import from vllm._C_stable_libtorch with %r", e) |
| 60 | |
| 61 | try: |
| 62 | import vllm._rocm_C # noqa: F401 |
| 63 | except ImportError as e: |
| 64 | logger.warning("Failed to import from vllm._rocm_C with %r", e) |
| 65 | |
| 66 | # Models not supported by ROCm. |
| 67 | _ROCM_UNSUPPORTED_MODELS: list[str] = [] |
| 68 | |
| 69 | # Models partially supported by ROCm. |
| 70 | # Architecture -> Reason. |
| 71 | _ROCM_PARTIALLY_SUPPORTED_MODELS: dict[str, str] = {} |
| 72 | _ROCM_DEVICE_ID_NAME_MAP: dict[str, str] = { |
| 73 | "0x74a0": "AMD_Instinct_MI300A", |
| 74 | "0x74a1": "AMD_Instinct_MI300X", |
| 75 | "0x74b5": "AMD_Instinct_MI300X", # MI300X VF |
| 76 | "0x74a2": "AMD_Instinct_MI308X", |
| 77 | "0x74a5": "AMD_Instinct_MI325X", |
| 78 | "0x74b9": "AMD_Instinct_MI325X", # MI325X VF |
| 79 | "0x74a9": "AMD_Instinct_MI300X_HF", |
| 80 | "0x74bd": "AMD_Instinct_MI300X_HF", |
| 81 | "0x744c": "AMD_Radeon_RX7900XTX", |
| 82 | # RDNA 3.5 APUs (Strix Point / Strix Halo) |
| 83 | "0x150e": "AMD_Radeon_890M", # gfx1150, Strix Point |
| 84 | "0x1586": "AMD_Radeon_8060S", # gfx1151, Strix Halo |
| 85 | # RDNA 4 discrete (Navi 48) |
| 86 | "0x7550": "AMD_Radeon_RX9070XT", # gfx1201 |
| 87 | "0x7551": "AMD_Radeon_R9700", # gfx1201 |
| 88 | } |
| 89 | |
| 90 | |
| 91 | @lru_cache(maxsize=8) |
| 92 | def _rocm_device_count_stateless(cuda_visible_devices: str | None = None) -> int: |
| 93 | """Get number of ROCm devices, caching based on the value of CUDA_VISIBLE_DEVICES |
| 94 | at the time of call. |
| 95 | |
| 96 | This should be used instead of torch.accelerator.device_count() unless |
| 97 | CUDA_VISIBLE_DEVICES has already been set to the desired value. |
| 98 | |
| 99 | # This can be removed and simply replaced with torch.cuda.get_device_count |
| 100 | # after https://github.com/pytorch/pytorch/pull/122815 is released.""" |
| 101 | # Note: cuda_visible_devices is not used, but we keep it as an argument for |
| 102 | # LRU Cache purposes. |
| 103 | |
| 104 | # Code below is based on |
| 105 | # https://github.com/pytorch/pytorch/blob/ |
| 106 | # c1cd946818442aca8c7f812b16d187ce1586c3bc/ |
| 107 | # torch/cuda/__init__.py#L831C1-L831C17 |
| 108 | import torch.cuda |
| 109 | |
| 110 | if not torch.cuda._is_compiled(): |
| 111 | return 0 |
| 112 | # ROCm uses amdsmi instead of nvml for stateless device count |
| 113 | # This requires a sufficiently modern version of Torch 2.4.0 |
| 114 | raw_count = ( |
| 115 | torch.cuda._device_count_amdsmi() |
| 116 | if (hasattr(torch.cuda, "_device_count_amdsmi")) |
| 117 | else -1 |
| 118 | ) |
| 119 | r = torch._C._cuda_getDeviceCount() if raw_count < 0 else raw_count |
| 120 | return r |
| 121 | |
| 122 | |
| 123 | @cache |
| 124 | def _get_wsl_kernel_version() -> tuple[int, ...] | None: |
| 125 | try: |
| 126 | release = platform.uname().release |
| 127 | parts = release.split("-")[0].split(".") |
| 128 | return tuple(int(part) for part in parts[:3]) |
| 129 | except (TypeError, ValueError): |
| 130 | return None |
| 131 | |
| 132 | |
| 133 | def _sync_hip_cuda_env_vars(): |
| 134 | """Ensure HIP_VISIBLE_DEVICES and CUDA_VISIBLE_DEVICES are consistent. |
| 135 | Treats empty string as unset. Raises on genuine conflicts.""" |
| 136 | hip_val = os.environ.get("HIP_VISIBLE_DEVICES") or None |
| 137 | cuda_val = os.environ.get("CUDA_VISIBLE_DEVICES") or None |
| 138 | |
| 139 | if hip_val is not None and cuda_val is not None: |
| 140 | if hip_val != cuda_val: |
| 141 | raise ValueError( |
| 142 | f"Inconsistent GPU visibility env vars: " |
| 143 | f"HIP_VISIBLE_DEVICES='{hip_val}' vs " |
| 144 | f"CUDA_VISIBLE_DEVICES='{cuda_val}'. " |
| 145 | f"Please set only one, or ensure they match." |
| 146 | ) |
| 147 | elif hip_val is not None: |
| 148 | os.environ["CUDA_VISIBLE_DEVICES"] = hip_val |
| 149 | |
| 150 | |
| 151 | # Sync at import time - catches misconfigurations from process start. |
| 152 | _sync_hip_cuda_env_vars() |
| 153 | |
| 154 | |
| 155 | # AMDSMI utils |
| 156 | # Note that NVML is not affected by `{CUDA/HIP}_VISIBLE_DEVICES`, |
| 157 | # all the related functions work on real physical device ids. |
| 158 | # the major benefit of using AMDSMI is that it will not initialize CUDA |
| 159 | |
| 160 | |
| 161 | def with_amdsmi_context(fn): |
| 162 | @wraps(fn) |
| 163 | def wrapper(*args, **kwargs): |
| 164 | amdsmi_init() |
| 165 | try: |
| 166 | return fn(*args, **kwargs) |
| 167 | finally: |
| 168 | amdsmi_shut_down() |
| 169 | |
| 170 | return wrapper |
| 171 | |
| 172 | |
| 173 | @with_amdsmi_context |
| 174 | def _query_gcn_arch_from_amdsmi() -> str: |
| 175 | """Query GCN arch from amdsmi. Raises if not available.""" |
| 176 | handles = amdsmi_get_processor_handles() |
| 177 | if handles: |
| 178 | asic_info = amdsmi_get_gpu_asic_info(handles[0]) |
| 179 | # Use target_graphics_version which contains the gfx name |
| 180 | # e.g., 'gfx942' for MI300X/MI325X |
| 181 | target_gfx = asic_info.get("target_graphics_version", "") |
| 182 | if target_gfx: |
| 183 | return target_gfx |
| 184 | raise RuntimeError("amdsmi did not return valid GCN arch") |
| 185 | |
| 186 | |
| 187 | @with_amdsmi_context |
| 188 | def _query_total_memory_from_amdsmi(physical_device_id: int) -> int: |
| 189 | """Query total VRAM (bytes) from amdsmi. Raises if not available.""" |
| 190 | handles = amdsmi_get_processor_handles() |
| 191 | handle = handles[physical_device_id] |
| 192 | return amdsmi_get_gpu_memory_total(handle, AmdSmiMemoryType.VRAM) |
| 193 | |
| 194 | |
| 195 | def _get_gcn_arch() -> str: |
| 196 | """Get GCN arch via amdsmi (no CUDA init), fallback to torch.cuda. |
| 197 | Called once at module level; result stored in _GCN_ARCH. |
| 198 | """ |
| 199 | try: |
| 200 | return _query_gcn_arch_from_amdsmi() |
| 201 | except Exception as e: |
| 202 | logger.debug("Failed to get GCN arch via amdsmi: %s", e) |
| 203 | # Ultimate fallback: use torch.cuda (will initialize CUDA) |
| 204 | return torch.cuda.get_device_properties("cuda").gcnArchName |
| 205 | |
| 206 | |
| 207 | # Resolve once at module load. Uses amdsmi (no CUDA init) so Ray workers |
| 208 | # can still set CUDA_VISIBLE_DEVICES after import. |
| 209 | # These are plain Python bools — fully torch.compile/Dynamo safe. |
| 210 | _GCN_ARCH = _get_gcn_arch() |
| 211 | |
| 212 | _ON_GFX1X = any(arch in _GCN_ARCH for arch in ["gfx11", "gfx12"]) |
| 213 | _ON_GFX11 = "gfx11" in _GCN_ARCH |
| 214 | _ON_GFX1100 = "gfx1100" in _GCN_ARCH |
| 215 | _ON_GFX1151 = "gfx1151" in _GCN_ARCH |
| 216 | _ON_GFX12X = any(arch in _GCN_ARCH for arch in ["gfx12"]) |
| 217 | _ON_MI3XX = any(arch in _GCN_ARCH for arch in ["gfx942", "gfx950"]) |
| 218 | _ON_GFX9 = any(arch in _GCN_ARCH for arch in ["gfx90a", "gfx942", "gfx950"]) |
| 219 | _ON_GFX90A = "gfx90a" in _GCN_ARCH |
| 220 | _ON_GFX942 = "gfx942" in _GCN_ARCH |
| 221 | _ON_GFX950 = "gfx950" in _GCN_ARCH |
| 222 | _ON_GFX1250 = "gfx1250" in _GCN_ARCH |
| 223 | |
| 224 | _ON_CDNA = any(arch in _GCN_ARCH for arch in ["gfx9", "gfx1250"]) |
| 225 | # RDNA = gfx11/gfx12 minus the CDNA-classified gfx1250. |
| 226 | _ON_RDNA = _ON_GFX1X and not _ON_CDNA |
| 227 | _ON_RDNA4 = any(arch in _GCN_ARCH for arch in ["gfx1200", "gfx1201"]) |
| 228 | |
| 229 | |
| 230 | def _capability_from_gcn_arch(gcn_arch: str) -> tuple[int, int] | None: |
| 231 | """Parse (major, minor) from a GCN arch string, mirroring how |
| 232 | HIP derives hipDeviceProp_t.major / .minor. |
| 233 | |
| 234 | Format: gfx<MAJOR><MINOR><STEPPING> |
| 235 | - 1-digit major (gfx9xx): "gfx" + M + m + stepping |
| 236 | - 2-digit major (gfx1xxx): "gfx" + MM + m + stepping |
| 237 | |
| 238 | Examples: |
| 239 | gfx90a -> (9, 0) gfx942 -> (9, 4) gfx950 -> (9, 5) |
| 240 | gfx1100 -> (11, 0) gfx1101 -> (11, 0) gfx1200 -> (12, 0) |
| 241 | |
| 242 | Returns None only when the string is not gfx-prefixed at all |
| 243 | (i.e. not a ROCm arch string). Raises on any string that looks |
| 244 | like a GCN arch but does not match a known layout. |
| 245 | |
| 246 | """ |
| 247 | m = re.match(r"gfx(\d+)", gcn_arch) |
| 248 | if not m: |
| 249 | # Not a gfx string at all — caller should fall back to torch.cuda |
| 250 | return None |
| 251 | |
| 252 | digits = m.group(1) |
| 253 | n = len(digits) |
| 254 | |
| 255 | if n < 2: |
| 256 | raise ValueError( |
| 257 | f"GCN arch '{gcn_arch}' has too few digits ({n}) after 'gfx' " |
| 258 | f"to derive a (major, minor) capability. " |
| 259 | f"Please file a vLLM issue with your GPU model." |
| 260 | ) |
| 261 | |
| 262 | if n in (2, 3): |
| 263 | # 1-digit major: gfx9 family |
| 264 | # len 2: major + minor (e.g. gfx90 from gfx90a) |
| 265 | # len 3: major + minor + step (e.g. gfx942) |
| 266 | major = int(digits[0]) |
| 267 | minor = int(digits[1]) |
| 268 | elif n == 4: |
| 269 | # 2-digit major: gfx10xx, gfx11xx, gfx12xx |
| 270 | # major(2) + minor(1) + stepping(1) |
| 271 | major = int(digits[:2]) |
| 272 | minor = int(digits[2]) |
| 273 | elif n >= 5: |
| 274 | raise ValueError( |
| 275 | f"GCN arch '{gcn_arch}' has {n} digits after 'gfx', which " |
| 276 | f"exceeds the known 4-digit layout (MMms). Cannot determine " |
| 277 | f"major/minor split unambiguously. " |
| 278 | f"Please file a vLLM issue with your GPU model." |
| 279 | ) |
| 280 | |
| 281 | if major < 9: |
| 282 | raise ValueError( |
| 283 | f"Parsed unknown ROCm architecture from GCN arch '{gcn_arch}': " |
| 284 | f"major={major}, minor={minor}. " |
| 285 | f"Major version < 9 is not expected for any supported AMD GPU. " |
| 286 | f"Please file a vLLM issue with your GPU model." |
| 287 | ) |
| 288 | |
| 289 | if major > 12: |
| 290 | raise ValueError( |
| 291 | f"Parsed unknown ROCm architecture from GCN arch '{gcn_arch}': " |
| 292 | f"major={major}, minor={minor}. " |
| 293 | f"Major version > 12 is beyond currently known AMD generations. " |
| 294 | f"Please file a vLLM issue with your GPU model so support " |
| 295 | f"can be added." |
| 296 | ) |
| 297 | |
| 298 | return (major, minor) |
| 299 | |
| 300 | |
| 301 | def on_gfx1x() -> bool: |
| 302 | return _ON_GFX1X and not _ON_CDNA |
| 303 | |
| 304 | |
| 305 | def on_gfx11() -> bool: |
| 306 | return _ON_GFX11 |
| 307 | |
| 308 | |
| 309 | def on_gfx1100() -> bool: |
| 310 | return _ON_GFX1100 |
| 311 | |
| 312 | |
| 313 | def on_gfx1151() -> bool: |
| 314 | return _ON_GFX1151 |
| 315 | |
| 316 | |
| 317 | def on_gfx12x() -> bool: |
| 318 | return _ON_GFX12X and not _ON_CDNA |
| 319 | |
| 320 | |
| 321 | def on_gfx1250() -> bool: |
| 322 | return _ON_GFX1250 |
| 323 | |
| 324 | |
| 325 | def on_rdna4() -> bool: |
| 326 | return _ON_RDNA4 |
| 327 | |
| 328 | |
| 329 | def on_mi3xx() -> bool: |
| 330 | return _ON_MI3XX |
| 331 | |
| 332 | |
| 333 | def on_gfx9() -> bool: |
| 334 | return _ON_GFX9 |
| 335 | |
| 336 | |
| 337 | def on_gfx90a() -> bool: |
| 338 | return _ON_GFX90A |
| 339 | |
| 340 | |
| 341 | def on_gfx942() -> bool: |
| 342 | return _ON_GFX942 |
| 343 | |
| 344 | |
| 345 | def on_gfx950() -> bool: |
| 346 | return _ON_GFX950 |
| 347 | |
| 348 | |
| 349 | def on_cdna() -> bool: |
| 350 | return _ON_CDNA |
| 351 | |
| 352 | |
| 353 | def on_rdna() -> bool: |
| 354 | return _ON_RDNA |
| 355 | |
| 356 | |
| 357 | def get_cdna_version() -> int: |
| 358 | if on_gfx90a(): |
| 359 | return 2 |
| 360 | if on_gfx942(): |
| 361 | return 3 |
| 362 | if on_gfx950(): |
| 363 | return 4 |
| 364 | if on_gfx1250(): |
| 365 | return 5 |
| 366 | return 0 |
| 367 | |
| 368 | |
| 369 | @cache |
| 370 | def get_rocm_version() -> tuple[int, ...] | None: |
| 371 | """Return the installed ROCm release as (major, minor, patch), or None.""" |
| 372 | # ROCm 10+ ships as the `rocm` pip SDK; older releases install to /opt/rocm. |
| 373 | try: |
| 374 | version = importlib.metadata.version("rocm") |
| 375 | except importlib.metadata.PackageNotFoundError: |
| 376 | rocm_path = os.environ.get("ROCM_PATH", "/opt/rocm") |
| 377 | try: |
| 378 | with open(os.path.join(rocm_path, ".info", "version")) as f: |
| 379 | version = f.read() |
| 380 | except OSError: |
| 381 | return None |
| 382 | match = re.match(r"\s*(\d+)\.(\d+)(?:\.(\d+))?", version) |
| 383 | if match is None: |
| 384 | return None |
| 385 | return tuple(int(part) for part in match.groups() if part is not None) |
| 386 | |
| 387 | |
| 388 | # Enable HIP online tuning early, before hipBLASLt initializes. |
| 389 | # Turn on hipBLASLt online tuning if use AITER hipBLASLt GEMM. |
| 390 | if ( |
| 391 | envs.VLLM_ROCM_USE_AITER |
| 392 | and envs.VLLM_ROCM_USE_AITER_LINEAR |
| 393 | and envs.VLLM_ROCM_USE_AITER_LINEAR_HIPBMM |
| 394 | and get_cdna_version() > 2 |
| 395 | ): |
| 396 | os.environ["HIP_ONLINE_TUNING"] = "1" |
| 397 | |
| 398 | |
| 399 | @cache |
| 400 | def use_rocm_custom_paged_attention( |
| 401 | qtype: torch.dtype, |
| 402 | head_size: int, |
| 403 | block_size: int, |
| 404 | gqa_ratio: int, |
| 405 | max_seq_len: int, |
| 406 | sliding_window: int, |
| 407 | kv_cache_dtype: str, |
| 408 | alibi_slopes: torch.Tensor | None = None, |
| 409 | sinks: torch.Tensor | None = None, |
| 410 | ) -> bool: |
| 411 | # custom paged attn always supported on V0. On V1, requires sliding window |
| 412 | # disabled due to observed numerical discrepancy. |
| 413 | if on_cdna(): |
| 414 | return ( |
| 415 | (sliding_window == 0 or sliding_window == (-1, -1)) |
| 416 | and (qtype == torch.half or qtype == torch.bfloat16) |
| 417 | and (head_size == 64 or head_size == 128) |
| 418 | and (block_size == 16 or block_size == 32) |
| 419 | and (gqa_ratio >= 1 and gqa_ratio <= 16) |
| 420 | and max_seq_len <= 128 * 1024 |
| 421 | and sinks is None |
| 422 | ) |
| 423 | |
| 424 | else: |
| 425 | return ( |
| 426 | _ON_GFX1X |
| 427 | and (sliding_window == 0 or sliding_window == (-1, -1)) |
| 428 | and (qtype == torch.half or qtype == torch.bfloat16) |
| 429 | and head_size == 128 |
| 430 | and block_size == 16 |
| 431 | and (gqa_ratio >= 3 and gqa_ratio <= 16) |
| 432 | and max_seq_len <= 128 * 1024 |
| 433 | and alibi_slopes is None |
| 434 | and kv_cache_dtype == "auto" |
| 435 | and sinks is None |
| 436 | ) |
| 437 | |
| 438 | |
| 439 | @cache |
| 440 | def flash_attn_triton_available() -> bool: |
| 441 | if not on_gfx1x(): |
| 442 | return False |
| 443 | try: |
| 444 | from importlib.util import find_spec |
| 445 | |
| 446 | # Locate the Triton-AMD kernels. Older ROCm/flash-attention (pre |
| 447 | # 2026-03) shipped them as the flash_attn.flash_attn_triton_amd |
| 448 | # subpackage. The main_perf migration commit 3f94643 moved them |
| 449 | # into aiter at aiter.ops.triton._triton_kernels.flash_attn_triton_amd, |
| 450 | # so accept either location. |
| 451 | def _has_spec(name: str) -> bool: |
| 452 | try: |
| 453 | return find_spec(name) is not None |
| 454 | except (ImportError, ValueError): |
| 455 | return False |
| 456 | |
| 457 | if not ( |
| 458 | _has_spec("flash_attn.flash_attn_triton_amd") |
| 459 | or _has_spec("aiter.ops.triton._triton_kernels.flash_attn_triton_amd") |
| 460 | ): |
| 461 | return False |
| 462 | if os.environ.get("FLASH_ATTENTION_TRITON_AMD_ENABLE") != "TRUE": |
| 463 | logger.info_once( |
| 464 | "Set FLASH_ATTENTION_TRITON_AMD_ENABLE=TRUE to enable " |
| 465 | "Flash Attention Triton backend on RDNA." |
| 466 | ) |
| 467 | return False |
| 468 | return True |
| 469 | except ImportError: |
| 470 | return False |
| 471 | |
| 472 | |
| 473 | def _get_backend_priorities( |
| 474 | use_mla: bool, |
| 475 | use_sparse: bool, |
| 476 | use_kv_connector: bool = False, |
| 477 | ) -> list[AttentionBackendEnum]: |
| 478 | from vllm._aiter_ops import is_aiter_found_and_supported, rocm_aiter_ops |
| 479 | |
| 480 | if use_sparse: |
| 481 | return [AttentionBackendEnum.ROCM_AITER_MLA_SPARSE] |
| 482 | |
| 483 | if use_mla: |
| 484 | if rocm_aiter_ops.is_mla_enabled(): |
| 485 | return [ |
| 486 | AttentionBackendEnum.ROCM_AITER_MLA, |
| 487 | AttentionBackendEnum.TRITON_MLA, |
| 488 | AttentionBackendEnum.ROCM_AITER_TRITON_MLA, |
| 489 | ] |
| 490 | else: |
| 491 | return [ |
| 492 | AttentionBackendEnum.TRITON_MLA, |
| 493 | ] |
| 494 | |
| 495 | backends = [] |
| 496 | # Keep ROCM_ATTN disabled for KV connectors until connector transfer |
| 497 | # semantics are validated for its asymmetric native K/V cache views. |
| 498 | if not use_kv_connector: |
| 499 | backends.append(AttentionBackendEnum.ROCM_ATTN) |
| 500 | if rocm_aiter_ops.is_mha_enabled(): |
| 501 | backends.append(AttentionBackendEnum.ROCM_AITER_FA) |
| 502 | if is_aiter_found_and_supported(): |
| 503 | backends.append(AttentionBackendEnum.ROCM_AITER_UNIFIED_ATTN) |
| 504 | elif rocm_aiter_ops.is_rdna_aiter_enabled(): |
| 505 | backends.insert(0, AttentionBackendEnum.ROCM_AITER_UNIFIED_ATTN) |
| 506 | backends.append(AttentionBackendEnum.TRITON_ATTN) |
| 507 | backends.append(AttentionBackendEnum.TURBOQUANT) |
| 508 | |
| 509 | return backends |
| 510 | |
| 511 | |
| 512 | def _uses_turboquant(vllm_config: "VllmConfig | None") -> bool: |
| 513 | """Whether the run's KV cache dtype is one of the turboquant_* presets.""" |
| 514 | cache_config = getattr(vllm_config, "cache_config", None) |
| 515 | return cache_config is not None and str(cache_config.cache_dtype).startswith( |
| 516 | "turboquant_" |
| 517 | ) |
| 518 | |
| 519 | |
| 520 | def _shares_layout_with_turboquant(backend_class: type["AttentionBackend"]) -> bool: |
| 521 | """Whether this backend reads a KV cache layout TURBOQUANT reads too. |
| 522 | |
| 523 | A turboquant_* run keeps its boundary layers at the native dtype (see |
| 524 | TurboQuantConfig.get_boundary_skip_layers), so those layers pick a backend of |
| 525 | their own while every other layer picks TURBOQUANT. One layout has to serve |
| 526 | the whole worker, and get_supported_kv_cache_layouts() turns an empty |
| 527 | intersection into a hard error at engine startup. |
| 528 | """ |
| 529 | layouts = backend_class.supported_kv_cache_layouts() |
| 530 | turboquant_layouts = ( |
| 531 | AttentionBackendEnum.TURBOQUANT.get_class().supported_kv_cache_layouts() |
| 532 | ) |
| 533 | if layouts is None or turboquant_layouts is None: |
| 534 | return True |
| 535 | return not set(layouts).isdisjoint(turboquant_layouts) |
| 536 | |
| 537 | |
| 538 | def _get_invalid_reasons( |
| 539 | backend_class: type["AttentionBackend"], |
| 540 | device_capability: DeviceCapability, |
| 541 | attn_selector_config: "AttentionSelectorConfig", |
| 542 | *, |
| 543 | is_turboquant_run: bool, |
| 544 | ) -> list[str]: |
| 545 | """Why this backend cannot serve the layer, empty when it can.""" |
| 546 | invalid_reasons = backend_class.validate_configuration( |
| 547 | device_capability=device_capability, |
| 548 | **attn_selector_config._asdict(), |
| 549 | ) |
| 550 | if ( |
| 551 | not invalid_reasons |
| 552 | and is_turboquant_run |
| 553 | and not _shares_layout_with_turboquant(backend_class) |
| 554 | ): |
| 555 | invalid_reasons = [_TURBOQUANT_LAYOUT_REASON] |
| 556 | return invalid_reasons |
| 557 | |
| 558 | |
| 559 | class RocmPlatform(Platform): |
| 560 | _enum = PlatformEnum.ROCM |
| 561 | device_name: str = "rocm" |
| 562 | device_type: str = "cuda" |
| 563 | dispatch_key: str = "CUDA" |
| 564 | ray_device_key: str = "GPU" |
| 565 | dist_backend: str = "nccl" |
| 566 | # rocm shares the same device control env var as CUDA |
| 567 | device_control_env_var: str = "CUDA_VISIBLE_DEVICES" |
| 568 | # Set in pre_register_and_update, so it exists only on the driver; Ray |
| 569 | # workers are separate processes and copy env vars by allowlist. |
| 570 | additional_env_vars: list[str] = ["GPU_PINNED_MIN_XFER_SIZE"] |
| 571 | ray_noset_device_env_vars: list[str] = [ |
| 572 | "RAY_EXPERIMENTAL_NOSET_HIP_VISIBLE_DEVICES", |
| 573 | "RAY_EXPERIMENTAL_NOSET_CUDA_VISIBLE_DEVICES", |
| 574 | "RAY_EXPERIMENTAL_NOSET_ROCR_VISIBLE_DEVICES", |
| 575 | ] |
| 576 | |
| 577 | supported_quantization: list[str] = [ |
| 578 | "awq", |
| 579 | "auto_awq", |
| 580 | "awq_marlin", # will be overwritten with awq |
| 581 | "gptq", |
| 582 | "auto_gptq", |
| 583 | "fp8", |
| 584 | "deepseek_v4_fp8", |
| 585 | "compressed-tensors", |
| 586 | "fbgemm_fp8", |
| 587 | "inc", |
| 588 | "quark", |
| 589 | "mxfp4", |
| 590 | "mxfp8", |
| 591 | "torchao", |
| 592 | "modelopt", |
| 593 | "modelopt_fp4", |
| 594 | "modelopt_mxfp8", |
| 595 | "modelopt_mixed", |
| 596 | "fp8_per_tensor", |
| 597 | "fp8_per_block", |
| 598 | "fp8_per_channel", |
| 599 | "online", |
| 600 | "gpt_oss_mxfp4", |
| 601 | ] |
| 602 | |
| 603 | @classmethod |
| 604 | def import_kernels(cls) -> None: |
| 605 | """Import ROCm-specific kernels.""" |
| 606 | super().import_kernels() |
| 607 | |
| 608 | import contextlib |
| 609 | |
| 610 | # Import ROCm-specific extension |
| 611 | with contextlib.suppress(ImportError): |
| 612 | import vllm._rocm_C # noqa: F401 |
| 613 | |
| 614 | @classmethod |
| 615 | def check_runner_kv_caches_multi_layer(cls) -> None: |
| 616 | pass |
| 617 | |
| 618 | @classmethod |
| 619 | def is_pin_memory_available(cls) -> bool: |
| 620 | if in_wsl(): |
| 621 | version = _get_wsl_kernel_version() |
| 622 | if version is None or version < (4, 19, 121): |
| 623 | # warning_once() causes a circular import on WSL, see #48397. |
| 624 | logger.warning( |
| 625 | "Using 'pin_memory=False' as WSL is detected and the " |
| 626 | "WSL2 kernel version is below 4.19.121. This may slow " |
| 627 | "down performance. Please run `wsl --update`." |
| 628 | ) |
| 629 | return False |
| 630 | |
| 631 | return True |
| 632 | |
| 633 | @classmethod |
| 634 | def get_valid_backends( |
| 635 | cls, |
| 636 | device_capability: DeviceCapability, |
| 637 | attn_selector_config: "AttentionSelectorConfig", |
| 638 | num_heads: int | None = None, |
| 639 | ) -> tuple[ |
| 640 | list[tuple["AttentionBackendEnum", int]], |
| 641 | dict["AttentionBackendEnum", list[str]], |
| 642 | ]: |
| 643 | valid_backends_priorities = [] |
| 644 | invalid_reasons = {} |
| 645 | |
| 646 | backend_priorities = _get_backend_priorities( |
| 647 | attn_selector_config.use_mla, |
| 648 | attn_selector_config.use_sparse, |
| 649 | attn_selector_config.use_kv_connector, |
| 650 | ) |
| 651 | from vllm.config import get_current_vllm_config_or_none |
| 652 | |
| 653 | vllm_config = get_current_vllm_config_or_none() |
| 654 | is_encoder_decoder = ( |
| 655 | getattr(getattr(vllm_config, "model_config", None), "attn_type", None) |
| 656 | == "encoder_decoder" |
| 657 | ) |
| 658 | # ROCM_ATTN still uses a legacy attention layout (KV is the outer |
| 659 | # dimension) that is incompatible with the encoder backend layouts. The |
| 660 | # encoder and decoder need the layouts to match. This is currently |
| 661 | # enforced implicitly. |
| 662 | # TODO: Make this explicit in the selector in a future PR. |
| 663 | if is_encoder_decoder and AttentionBackendEnum.ROCM_ATTN in backend_priorities: |
| 664 | backend_priorities.remove(AttentionBackendEnum.ROCM_ATTN) |
| 665 | is_turboquant_run = _uses_turboquant(vllm_config) |
| 666 | for priority, backend in enumerate(backend_priorities): |
| 667 | try: |
| 668 | invalid_reasons_i = _get_invalid_reasons( |
| 669 | backend.get_class(), |
| 670 | device_capability, |
| 671 | attn_selector_config, |
| 672 | is_turboquant_run=is_turboquant_run, |
| 673 | ) |
| 674 | except ImportError: |
| 675 | invalid_reasons_i = ["ImportError"] |
| 676 | if invalid_reasons_i: |
| 677 | invalid_reasons[backend] = invalid_reasons_i |
| 678 | else: |
| 679 | valid_backends_priorities.append((backend, priority)) |
| 680 | |
| 681 | return valid_backends_priorities, invalid_reasons |
| 682 | |
| 683 | @classmethod |
| 684 | def get_attn_backend_cls( |
| 685 | cls, |
| 686 | selected_backend: "AttentionBackendEnum", |
| 687 | attn_selector_config: "AttentionSelectorConfig", |
| 688 | num_heads: int | None = None, |
| 689 | ) -> str: |
| 690 | device_capability = cls.get_device_capability() |
| 691 | assert device_capability is not None |
| 692 | |
| 693 | # First try checking just the selected backend, if there is one. |
| 694 | if selected_backend is not None: |
| 695 | # Keep lazy: vllm.config imports current_platform during initialization. |
| 696 | from vllm.config import get_current_vllm_config_or_none |
| 697 | |
| 698 | is_turboquant_run = _uses_turboquant(get_current_vllm_config_or_none()) |
| 699 | try: |
| 700 | sel_invalid_reasons = _get_invalid_reasons( |
| 701 | selected_backend.get_class(), |
| 702 | device_capability, |
| 703 | attn_selector_config, |
| 704 | is_turboquant_run=is_turboquant_run, |
| 705 | ) |
| 706 | except ImportError: |
| 707 | sel_invalid_reasons = ["ImportError"] |
| 708 | if not sel_invalid_reasons: |
| 709 | logger.info_once( |
| 710 | "Using %s backend (selected via --attention-backend).", |
| 711 | selected_backend.name, |
| 712 | ) |
| 713 | return selected_backend.get_path() |
| 714 | # Only tolerate the mismatch when turboquant is in play: boundary |
| 715 | # layers keep the native dtype while every other layer needs |
| 716 | # TURBOQUANT, so no single --attention-backend can serve every layer. |
| 717 | # For any other dtype the selection is genuinely invalid -> fail loud. |
| 718 | kv_dtype = attn_selector_config.kv_cache_dtype |
| 719 | layer_is_turboquant = kv_dtype is not None and str(kv_dtype).startswith( |
| 720 | "turboquant" |
| 721 | ) |
| 722 | is_turboquant_fallback = ( |
| 723 | is_turboquant_run or layer_is_turboquant |
| 724 | ) and sel_invalid_reasons in ( |
| 725 | [_KV_CACHE_DTYPE_REASON], |
| 726 | [_TURBOQUANT_LAYOUT_REASON], |
| 727 | ) |
| 728 | if not is_turboquant_fallback: |
| 729 | raise ValueError( |
| 730 | f"Selected backend {selected_backend} is not valid for " |
| 731 | f"this configuration. Reason: {sel_invalid_reasons}" |
| 732 | ) |
| 733 | # NOTE: pass a str (not the list) -- info_once hashes its args. |
| 734 | logger.info_once( |
| 735 | "Selected backend %s is incompatible with this layer (%s) of " |
| 736 | "the turboquant run; using the auto-selected per-layer backend. " |
| 737 | "Reason: %s", |
| 738 | selected_backend.name, |
| 739 | attn_selector_config.attn_type, |
| 740 | str(sel_invalid_reasons), |
| 741 | ) |
| 742 | |
| 743 | # No selected backend or the selected backend is invalid, |
| 744 | # so we try finding a valid backend. |
| 745 | valid_backends_priorities, invalid_reasons = cls.get_valid_backends( |
| 746 | device_capability=device_capability, |
| 747 | attn_selector_config=attn_selector_config, |
| 748 | num_heads=num_heads, |
| 749 | ) |
| 750 | reasons_str = ( |
| 751 | "{" |
| 752 | + ", ".join( |
| 753 | f"{backend.name}: [{', '.join(reasons)}]" |
| 754 | for backend, reasons in invalid_reasons.items() |
| 755 | ) |
| 756 | + "}" |
| 757 | ) |
| 758 | config_str = attn_selector_config.__repr__() |
| 759 | logger.debug_once( |
| 760 | f"Some attention backends are not valid for {cls.device_name} with " |
| 761 | f"{config_str}. Reasons: {reasons_str}." |
| 762 | ) |
| 763 | if len(valid_backends_priorities) == 0: |
| 764 | # If a backend rejected the requested kv-cache dtype, list the |
| 765 | # dtypes it does accept so the limitation is discoverable. |
| 766 | supported = sorted( |
| 767 | { |
| 768 | dt |
| 769 | for backend, reasons in invalid_reasons.items() |
| 770 | if any("kv_cache_dtype" in r for r in reasons) |
| 771 | for dt in backend.get_class().supported_kv_cache_dtypes |
| 772 | } |
| 773 | ) |
| 774 | hint = ( |
| 775 | f" Supported kv_cache_dtype values: {', '.join(supported)}." |
| 776 | if supported |
| 777 | else "" |
| 778 | ) |
| 779 | raise ValueError( |
| 780 | f"No valid attention backend found for {cls.device_name} " |
| 781 | f"with {config_str}. Reasons: {reasons_str}.{hint}" |
| 782 | ) |
| 783 | |
| 784 | # We have found some valid backends. Select the one with the |
| 785 | # highest priority. |
| 786 | sorted_indices = sorted( |
| 787 | range(len(valid_backends_priorities)), |
| 788 | key=lambda i: valid_backends_priorities[i][1], |
| 789 | ) |
| 790 | selected_index = sorted_indices[0] |
| 791 | selected_backend = valid_backends_priorities[selected_index][0] |
| 792 | valid_str = ( |
| 793 | "[" + ", ".join(f"'{b[0].name}'" for b in valid_backends_priorities) + "]" |
| 794 | ) |
| 795 | if invalid_reasons: |
| 796 | rejected_str = ", ".join(b.name for b in invalid_reasons) |
| 797 | logger.info( |
| 798 | "Found incompatible backend(s) [%s] with %s. " |
| 799 | "Overriding with %s out of potential backends: %s.", |
| 800 | rejected_str, |
| 801 | attn_selector_config.attn_type, |
| 802 | selected_backend.name, |
| 803 | valid_str, |
| 804 | ) |
| 805 | else: |
| 806 | logger.info_once( |
| 807 | "Using %s backend out of potential backends: %s.", |
| 808 | selected_backend.name, |
| 809 | valid_str, |
| 810 | ) |
| 811 | |
| 812 | return selected_backend.get_path() |
| 813 | |
| 814 | @classmethod |
| 815 | def get_supported_vit_attn_backends(cls) -> list["AttentionBackendEnum"]: |
| 816 | return [ |
| 817 | AttentionBackendEnum.FLASH_ATTN, |
| 818 | AttentionBackendEnum.ROCM_AITER_FA, |
| 819 | AttentionBackendEnum.TRITON_ATTN, |
| 820 | AttentionBackendEnum.TORCH_SDPA, |
| 821 | ] |
| 822 | |
| 823 | @classmethod |
| 824 | def get_vit_attn_backend( |
| 825 | cls, |
| 826 | head_size: int, |
| 827 | dtype: torch.dtype, |
| 828 | backend: "AttentionBackendEnum | None" = None, |
| 829 | ) -> "AttentionBackendEnum": |
| 830 | if backend is not None: |
| 831 | assert backend in cls.get_supported_vit_attn_backends(), ( |
| 832 | f"Backend {backend} is not supported for vit attention. " |
| 833 | f"Supported backends are: {cls.get_supported_vit_attn_backends()}" |
| 834 | ) |
| 835 | logger.info_once(f"Using backend {backend} for vit attention") |
| 836 | return backend |
| 837 | |
| 838 | from importlib.util import find_spec |
| 839 | |
| 840 | from vllm._aiter_ops import rocm_aiter_ops |
| 841 | |
| 842 | if rocm_aiter_ops.is_mha_enabled() and on_cdna(): |
| 843 | logger.info_once("Using AITER Flash Attention backend for ViT model.") |
| 844 | return AttentionBackendEnum.ROCM_AITER_FA |
| 845 | |
| 846 | if ( |
| 847 | on_cdna() |
| 848 | and find_spec("flash_attn") is not None |
| 849 | and (dtype == torch.float16 or dtype == torch.bfloat16) |
| 850 | ): |
| 851 | logger.info_once("Using Flash Attention backend for ViT model.") |
| 852 | return AttentionBackendEnum.FLASH_ATTN |
| 853 | |
| 854 | # RDNA3/RDNA4 (gfx11xx/gfx12xx): Use Flash Attention Triton backend |
| 855 | if ( |
| 856 | on_gfx1x() |
| 857 | and flash_attn_triton_available() |
| 858 | and (dtype == torch.float16 or dtype == torch.bfloat16) |
| 859 | ): |
| 860 | logger.info_once( |
| 861 | "Using Flash Attention (Triton backend) for ViT model on RDNA." |
| 862 | ) |
| 863 | return AttentionBackendEnum.FLASH_ATTN |
| 864 | |
| 865 | logger.info_once("Using Torch SDPA backend for ViT model.") |
| 866 | return AttentionBackendEnum.TORCH_SDPA |
| 867 | |
| 868 | @classmethod |
| 869 | def set_device(cls, device: torch.device) -> None: |
| 870 | """Set the device for the current platform.""" |
| 871 | torch.cuda.set_device(device) |
| 872 | |
| 873 | @classmethod |
| 874 | def manual_seed_all(cls, seed: int) -> None: |
| 875 | torch.cuda.manual_seed_all(seed) |
| 876 | |
| 877 | @classmethod |
| 878 | @lru_cache(maxsize=8) |
| 879 | def get_device_capability(cls, device_id: int = 0) -> DeviceCapability | None: |
| 880 | cap = _capability_from_gcn_arch(_GCN_ARCH) |
| 881 | if cap is not None: |
| 882 | return DeviceCapability(major=cap[0], minor=cap[1]) |
| 883 | |
| 884 | logger.warning_once( |
| 885 | "Could not derive device capability from GCN arch '%s', " |
| 886 | "falling back to torch.cuda (this will initialize CUDA).", |
| 887 | _GCN_ARCH, |
| 888 | ) |
| 889 | major, minor = torch.cuda.get_device_capability(device_id) |
| 890 | return DeviceCapability(major=major, minor=minor) |
| 891 | |
| 892 | @classmethod |
| 893 | @with_amdsmi_context |
| 894 | def is_fully_connected(cls, physical_device_ids: list[int]) -> bool: |
| 895 | """Query if the set of gpus are fully connected by xgmi (1 hop).""" |
| 896 | handles = [amdsmi_get_processor_handles()[i] for i in physical_device_ids] |
| 897 | for i, handle in enumerate(handles): |
| 898 | for j, peer_handle in enumerate(handles): |
| 899 | if i < j: |
| 900 | try: |
| 901 | link_type = amdsmi_topo_get_link_type(handle, peer_handle) |
| 902 | # type is 2 for XGMI |
| 903 | if link_type["hops"] != 1 or link_type["type"] != 2: |
| 904 | return False |
| 905 | except AmdSmiException as error: |
| 906 | logger.error("AMD 1 hop XGMI detection failed.", exc_info=error) |
| 907 | return False |
| 908 | return True |
| 909 | |
| 910 | @classmethod |
| 911 | @with_amdsmi_context |
| 912 | @lru_cache(maxsize=8) |
| 913 | def get_device_name(cls, device_id: int = 0) -> str: |
| 914 | physical_device_id = cls.device_id_to_physical_device_id(device_id) |
| 915 | handle = amdsmi_get_processor_handles()[physical_device_id] |
| 916 | asic_info = amdsmi_get_gpu_asic_info(handle) |
| 917 | asic_info_device_id: str = asic_info["device_id"] |
| 918 | if asic_info_device_id in _ROCM_DEVICE_ID_NAME_MAP: |
| 919 | return _ROCM_DEVICE_ID_NAME_MAP[asic_info_device_id] |
| 920 | return asic_info["market_name"] |
| 921 | |
| 922 | @classmethod |
| 923 | @with_amdsmi_context |
| 924 | def get_device_uuid(cls, device_id: int = 0) -> str: |
| 925 | try: |
| 926 | device = amdsmi_get_processor_handles()[device_id] |
| 927 | except AmdSmiException as error: |
| 928 | logger.error("amdsmi device query failed ", exc_info=error) |
| 929 | return "" |
| 930 | try: |
| 931 | device_uuid = amdsmi_get_gpu_device_uuid(device) |
| 932 | except AmdSmiException as error: |
| 933 | logger.error("amdsmi device uuid query failed ", exc_info=error) |
| 934 | return device_uuid |
| 935 | |
| 936 | @classmethod |
| 937 | def get_device_total_memory(cls, device_id: int = 0) -> int: |
| 938 | # Query total VRAM via amdsmi so we don't initialize a HIP context in |
| 939 | # the calling process. torch.cuda.get_device_properties() creates a |
| 940 | # HIP context, which makes vLLM fall back from `fork` to `spawn` for |
| 941 | # worker processes. Keeping this query context-free preserves `fork` |
| 942 | # where it is otherwise valid (e.g. out-of-tree models registered in |
| 943 | # the parent process). |
| 944 | try: |
| 945 | physical_device_id = cls.device_id_to_physical_device_id(device_id) |
| 946 | return _query_total_memory_from_amdsmi(physical_device_id) |
| 947 | except Exception as e: |
| 948 | logger.debug("Failed to get total memory via amdsmi: %s", e) |
| 949 | logger.warning_once( |
| 950 | "Failed to get total memory via amdsmi, falling back to " |
| 951 | "torch.cuda. This will initialize CUDA." |
| 952 | ) |
| 953 | return torch.cuda.get_device_properties(device_id).total_memory |
| 954 | |
| 955 | @classmethod |
| 956 | def pre_register_and_update( |
| 957 | cls, parser: "FlexibleArgumentParser | None" = None |
| 958 | ) -> None: |
| 959 | # Keep mmap'd weight pages on the HIP staging path: above this |
| 960 | # threshold the runtime registers the pageable source instead, and each |
| 961 | # registration's MMU notifier makes KFD suspend our queues. In KB, so |
| 962 | # 4 GiB. |
| 963 | os.environ.setdefault("GPU_PINNED_MIN_XFER_SIZE", str(4 * 1024 * 1024)) |
| 964 | |
| 965 | @classmethod |
| 966 | def apply_config_platform_defaults(cls, vllm_config: "VllmConfig") -> None: |
| 967 | from vllm._aiter_ops import rocm_aiter_ops |
| 968 | |
| 969 | compilation_config = vllm_config.compilation_config |
| 970 | use_aiter_fused_moe = rocm_aiter_ops.is_fused_moe_enabled() |
| 971 | use_aiter_fp8_linear = rocm_aiter_ops.is_linear_fp8_enabled() |
| 972 | use_aiter_fused_se = rocm_aiter_ops.is_fusion_moe_shared_experts_enabled() |
| 973 | |
| 974 | if use_aiter_fp8_linear and "-quant_fp8" not in compilation_config.custom_ops: |
| 975 | compilation_config.custom_ops.append("+quant_fp8") |
| 976 | |
| 977 | if use_aiter_fused_se and "-grouped_topk" in compilation_config.custom_ops: |
| 978 | logger.warning_once( |
| 979 | "VLLM_ROCM_USE_AITER_FUSION_SHARED_EXPERTS is enabled, which " |
| 980 | "requires the 'grouped_topk' custom op. Overriding the " |
| 981 | "user-provided '-grouped_topk'." |
| 982 | ) |
| 983 | compilation_config.custom_ops.remove("-grouped_topk") |
| 984 | # Ensure grouped_topk is always enabled when using AITER if |
| 985 | # its not disabled by user |
| 986 | if ( |
| 987 | use_aiter_fused_moe |
| 988 | and "+grouped_topk" not in compilation_config.custom_ops |
| 989 | and "-grouped_topk" not in compilation_config.custom_ops |
| 990 | ): |
| 991 | compilation_config.custom_ops.append("+grouped_topk") |
| 992 | |
| 993 | # Default dispatch to rocm's sparse_attn_indexer implementation |
| 994 | compilation_config.custom_ops.append("+sparse_attn_indexer") |
| 995 | |
| 996 | @classmethod |
| 997 | def check_and_update_config(cls, vllm_config: "VllmConfig") -> None: |
| 998 | from vllm.config.compilation import CUDAGraphMode |
| 999 | |
| 1000 | compilation_config = vllm_config.compilation_config |
| 1001 | parallel_config = vllm_config.parallel_config |
| 1002 | |
| 1003 | if ( |
| 1004 | compilation_config.cudagraph_mode.has_full_cudagraphs() |
| 1005 | and parallel_config.prefill_context_parallel_size > 1 |
| 1006 | ): |
| 1007 | # prefill context parallel do not support full cudagraphs |
| 1008 | logger.warning_once( |
| 1009 | "Prefill context parallel (PCP) is enabled, which is " |
| 1010 | "incompatible with full CUDA graphs. " |
| 1011 | "Overriding cudagraph_mode to PIECEWISE." |
| 1012 | ) |
| 1013 | compilation_config.cudagraph_mode = CUDAGraphMode.PIECEWISE |
| 1014 | |
| 1015 | if parallel_config.worker_cls == "auto": |
| 1016 | parallel_config.worker_cls = "vllm.v1.worker.gpu_worker.Worker" |
| 1017 | |
| 1018 | model_config = vllm_config.model_config |
| 1019 | scheduler_config = vllm_config.scheduler_config |
| 1020 | # Note: model_config may be None during testing |
| 1021 | if ( |
| 1022 | model_config is not None |
| 1023 | and model_config.is_mm_prefix_lm |
| 1024 | and scheduler_config.is_multimodal_model |
| 1025 | and not scheduler_config.disable_chunked_mm_input |
| 1026 | ): |
| 1027 | logger.warning_once( |
| 1028 | "Forcing --disable_chunked_mm_input for models " |
| 1029 | "with multimodal-bidirectional attention." |
| 1030 | ) |
| 1031 | scheduler_config.disable_chunked_mm_input = True |
| 1032 | |
| 1033 | @classmethod |
| 1034 | def verify_model_arch(cls, model_arch: str) -> None: |
| 1035 | if model_arch in _ROCM_UNSUPPORTED_MODELS: |
| 1036 | raise ValueError( |
| 1037 | f"Model architecture '{model_arch}' is not supported by ROCm for now." |
| 1038 | ) |
| 1039 | |
| 1040 | if model_arch in _ROCM_PARTIALLY_SUPPORTED_MODELS: |
| 1041 | msg = _ROCM_PARTIALLY_SUPPORTED_MODELS[model_arch] |
| 1042 | logger.warning( |
| 1043 | "Model architecture '%s' is partially supported by ROCm: %s", |
| 1044 | model_arch, |
| 1045 | msg, |
| 1046 | ) |
| 1047 | |
| 1048 | @classmethod |
| 1049 | def verify_quantization(cls, quant: str) -> None: |
| 1050 | super().verify_quantization(quant) |
| 1051 | if quant == "awq" and not envs.VLLM_USE_TRITON_AWQ: |
| 1052 | logger.warning( |
| 1053 | "Using AWQ quantization with ROCm, but VLLM_USE_TRITON_AWQ" |
| 1054 | " is not set, enabling VLLM_USE_TRITON_AWQ." |
| 1055 | ) |
| 1056 | os.environ["VLLM_USE_TRITON_AWQ"] = "1" |
| 1057 | |
| 1058 | @classmethod |
| 1059 | def get_punica_wrapper(cls) -> str: |
| 1060 | return "vllm.lora.punica_wrapper.punica_gpu.PunicaWrapperGPU" |
| 1061 | |
| 1062 | @classmethod |
| 1063 | def get_current_memory_usage( |
| 1064 | cls, device: torch.types.Device | None = None |
| 1065 | ) -> float: |
| 1066 | torch.cuda.empty_cache() |
| 1067 | torch.cuda.reset_peak_memory_stats(device) |
| 1068 | return torch.cuda.max_memory_allocated(device) |
| 1069 | |
| 1070 | @classmethod |
| 1071 | def get_device_communicator_cls(cls) -> str: |
| 1072 | return ( |
| 1073 | "vllm.distributed.device_communicators.cuda_communicator.CudaCommunicator" # noqa |
| 1074 | ) |
| 1075 | |
| 1076 | @classmethod |
| 1077 | def supports_mx(cls) -> bool: |
| 1078 | return any(gfx in _GCN_ARCH for gfx in ["gfx95", "gfx1250"]) |
| 1079 | |
| 1080 | @classmethod |
| 1081 | def supports_fp8(cls) -> bool: |
| 1082 | return on_cdna() or on_rdna4() |
| 1083 | |
| 1084 | @classmethod |
| 1085 | def is_fp8_fnuz(cls) -> bool: |
| 1086 | # only device 0 is checked, this assumes MI300 platforms are homogeneous |
| 1087 | return "gfx94" in _GCN_ARCH |
| 1088 | |
| 1089 | @classmethod |
| 1090 | def fp8_dtype(cls) -> torch.dtype: |
| 1091 | if cls.is_fp8_fnuz(): |
| 1092 | return torch.float8_e4m3fnuz |
| 1093 | else: |
| 1094 | return torch.float8_e4m3fn |
| 1095 | |
| 1096 | @classmethod |
| 1097 | def use_custom_allreduce(cls) -> bool: |
| 1098 | # We only enable custom allreduce for MI300 series |
| 1099 | return any(gfx in _GCN_ARCH for gfx in ["gfx94", "gfx95"]) |
| 1100 | |
| 1101 | @classmethod |
| 1102 | def opaque_attention_op(cls) -> bool: |
| 1103 | return True |
| 1104 | |
| 1105 | @classmethod |
| 1106 | def is_navi(cls) -> bool: |
| 1107 | return "gfx1" in _GCN_ARCH |
| 1108 | |
| 1109 | @classmethod |
| 1110 | def enable_multi_stream_overlap( |
| 1111 | cls, |
| 1112 | aux_stream_list: list[torch.cuda.Stream] | None, |
| 1113 | attn_metadata: object, |
| 1114 | ) -> bool: |
| 1115 | """ROCm multi-stream gates: streams and capture region. |
| 1116 | |
| 1117 | Dict metadata marks piecewise cudagraph, whose eager breaks rebuild |
| 1118 | the attention inputs on the owning stream. Forking side streams |
| 1119 | there would rely on runtime HIP event sync, which is unreliable in |
| 1120 | this overlap on ROCm (event waits can hang), so multi-stream only |
| 1121 | runs where the fork/join becomes static graph edges: inside capture, |
| 1122 | or with non-dict metadata (full cudagraph or the profile run), which |
| 1123 | has no eager breaks. |
| 1124 | """ |
| 1125 | return aux_stream_list is not None and ( |
| 1126 | torch.cuda.is_current_stream_capturing() |
| 1127 | or not isinstance(attn_metadata, dict) |
| 1128 | ) |
| 1129 | |
| 1130 | @classmethod |
| 1131 | def get_static_graph_wrapper_cls(cls) -> str: |
| 1132 | return "vllm.compilation.cuda_graph.CUDAGraphWrapper" |
| 1133 | |
| 1134 | @classmethod |
| 1135 | def stateless_init_device_torch_dist_pg( |
| 1136 | cls, |
| 1137 | backend: str, |
| 1138 | prefix_store: PrefixStore, |
| 1139 | group_rank: int, |
| 1140 | group_size: int, |
| 1141 | timeout: timedelta, |
| 1142 | ) -> ProcessGroup: |
| 1143 | assert is_nccl_available() |
| 1144 | pg: ProcessGroup = ProcessGroup( |
| 1145 | prefix_store, |
| 1146 | group_rank, |
| 1147 | group_size, |
| 1148 | ) |
| 1149 | from torch.distributed.distributed_c10d import ProcessGroupNCCL |
| 1150 | |
| 1151 | backend_options = ProcessGroupNCCL.Options() |
| 1152 | backend_options._timeout = timeout |
| 1153 | |
| 1154 | backend_class = ProcessGroupNCCL( |
| 1155 | prefix_store, group_rank, group_size, backend_options |
| 1156 | ) |
| 1157 | backend_type = ProcessGroup.BackendType.NCCL |
| 1158 | device = torch.device("cuda") |
| 1159 | pg._set_default_backend(backend_type) |
| 1160 | backend_class._set_sequence_number_for_group() |
| 1161 | |
| 1162 | pg._register_backend(device, backend_type, backend_class) |
| 1163 | return pg |
| 1164 | |
| 1165 | @classmethod |
| 1166 | def device_count(cls) -> int: |
| 1167 | return _rocm_device_count_stateless(getattr(envs, cls.device_control_env_var)) |
| 1168 | |
| 1169 | @classmethod |
| 1170 | def check_if_supports_dtype(cls, dtype: torch.dtype): |
| 1171 | if dtype == torch.bfloat16: # noqa: SIM102 |
| 1172 | if not cls.has_device_capability(80): |
| 1173 | capability = cls.get_device_capability() |
| 1174 | gpu_name = cls.get_device_name() |
| 1175 | |
| 1176 | if capability is None: |
| 1177 | compute_str = "does not have a compute capability" |
| 1178 | else: |
| 1179 | version_str = capability.as_version_str() |
| 1180 | compute_str = f"has compute capability {version_str}" |
| 1181 | |
| 1182 | raise ValueError( |
| 1183 | "Bfloat16 is only supported on GPUs " |
| 1184 | "with compute capability of at least 8.0. " |
| 1185 | f"Your {gpu_name} GPU {compute_str}. " |
| 1186 | "You can use float16 instead by explicitly setting the " |
| 1187 | "`dtype` flag in CLI, for example: --dtype=half." |
| 1188 | ) |
| 1189 | |
| 1190 | @classmethod |
| 1191 | def insert_blocks_to_device( |
| 1192 | cls, |
| 1193 | src_cache: torch.Tensor, |
| 1194 | dst_cache: torch.Tensor, |
| 1195 | src_block_indices: torch.Tensor, |
| 1196 | dst_block_indices: torch.Tensor, |
| 1197 | ) -> None: |
| 1198 | """Copy blocks from src_cache to dst_cache on GPU.""" |
| 1199 | _src_cache = src_cache[src_block_indices] |
| 1200 | dst_cache[dst_block_indices] = _src_cache.to(dst_cache.device) |
| 1201 | |
| 1202 | @classmethod |
| 1203 | def swap_out_blocks_to_host( |
| 1204 | cls, |
| 1205 | src_cache: torch.Tensor, |
| 1206 | dst_cache: torch.Tensor, |
| 1207 | src_block_indices: torch.Tensor, |
| 1208 | dst_block_indices: torch.Tensor, |
| 1209 | ) -> None: |
| 1210 | """Copy blocks from GPU to host (CPU).""" |
| 1211 | _src_cache = src_cache[src_block_indices] |
| 1212 | dst_cache[dst_block_indices] = _src_cache.cpu() |
| 1213 | |
| 1214 | @classmethod |
| 1215 | def support_hybrid_kv_cache(cls) -> bool: |
| 1216 | return True |
| 1217 | |
| 1218 | @classmethod |
| 1219 | def support_static_graph_mode(cls) -> bool: |
| 1220 | return True |
| 1221 | |
| 1222 | @classmethod |
| 1223 | def num_compute_units(cls, device_id: int = 0) -> int: |
| 1224 | return torch.cuda.get_device_properties(device_id).multi_processor_count |
| 1225 | |
| 1226 | @classmethod |
| 1227 | def use_custom_op_collectives(cls) -> bool: |
| 1228 | return True |
| 1229 | |
| 1230 | @classmethod |
| 1231 | def get_default_ir_op_priority( |
| 1232 | cls, vllm_config: "VllmConfig" |
| 1233 | ) -> "IrOpPriorityConfig": |
| 1234 | from vllm.config.compilation import CompilationMode, CUDAGraphMode |
| 1235 | from vllm.config.kernel import IrOpPriorityConfig |
| 1236 | |
| 1237 | # Native used by default when compiling, |
| 1238 | # use vllm_c kernels where available when no codegen |
| 1239 | # TODO(luka/TJ) use aiter, vllm_c, native by default on ROCm |
| 1240 | cc = vllm_config.compilation_config |
| 1241 | using_inductor = cc.backend == "inductor" and cc.mode != CompilationMode.NONE |
| 1242 | default = ["native"] if using_inductor else ["vllm_c", "native"] |
| 1243 | |
| 1244 | # Aiter rms norm perform best when CUDA Graph capture is enabled. |
| 1245 | # TODO(luka/TJ) remove env vars completely |
| 1246 | if ( |
| 1247 | cc.cudagraph_mode != CUDAGraphMode.NONE |
| 1248 | and envs.VLLM_ROCM_USE_AITER |
| 1249 | and envs.VLLM_ROCM_USE_AITER_RMSNORM |
| 1250 | and not on_rdna4() |
| 1251 | ): |
| 1252 | rms_norm = ["aiter"] + default |
| 1253 | else: |
| 1254 | rms_norm = default |
| 1255 | |
| 1256 | return IrOpPriorityConfig.with_default( |
| 1257 | default, |
| 1258 | rms_norm=rms_norm, |
| 1259 | fused_add_rms_norm=rms_norm, |
| 1260 | gelu_and_mul_sparse=["native"], |
| 1261 | ) |
| 1262 | |
| 1263 | @classmethod |
| 1264 | @with_amdsmi_context |
| 1265 | def get_all_device_numa_nodes(cls) -> list[int] | None: |
| 1266 | """Get NUMA nodes for all visible GPU devices.""" |
| 1267 | try: |
| 1268 | handles = amdsmi_get_processor_handles() |
| 1269 | numa_nodes = [] |
| 1270 | for device_id in range(cls.device_count()): |
| 1271 | physical_device_id = cls.device_id_to_physical_device_id(device_id) |
| 1272 | try: |
| 1273 | numa_node = amdsmi_topo_get_numa_node_number( |
| 1274 | handles[physical_device_id] |
| 1275 | ) |
| 1276 | except AmdSmiException as e: |
| 1277 | logger.warning( |
| 1278 | "Could not detect NUMA node for GPU %d, " |
| 1279 | "disabling automatic NUMA binding: %s", |
| 1280 | device_id, |
| 1281 | e, |
| 1282 | ) |
| 1283 | return None |
| 1284 | numa_nodes.append(numa_node) |
| 1285 | return numa_nodes |
| 1286 | except Exception as e: |
| 1287 | logger.warning("Failed to get NUMA nodes for GPUs: %s", e) |
| 1288 | return None |