A symbol scan over every tracked file found 123 top-level definitions in vllm/ with zero references, including references inside strings. All of them came from features this fork deleted earlier. - Delete dead attention helpers: MLADCPManager, split_prefill_chunks, the reshape helpers, compute_varlen_chunk_metadata, the LINEAR backend's unused BailingLinearAttentionBackend, and the sparse flash-attn entry points. - Delete the CPU and AMX op wrappers from _custom_ops.py: cpu_mla_decode, cpu_mla_extend, cpu_mla_get_scheduler_metadata, gather_mla_context_cache_cpu. - Delete dead quantization helpers: awq_pack, gptq_pack, gptq_quantize_weights, get_and_maybe_dequant_weights, marlin_quant_fp8_torch, the rand_marlin_weight_*_like pair, ocp_mx_weight_dtype_and_rows, get_fp8_block_weight_scale, input_to_float8. - Delete dead model helpers: resolve_visual_encoder_outputs, run_dp_sharded_vision_model, fast_topk, get_layer_index, is_pp_missing_parameter, skip_spec_layers, KimiK3MergedQKVGateLinear, selective_scan_fn, causal_conv1d_update_torch. - Delete dead utils: format_kib, format_mib, chunk_list, prev_power_of_2, split_host_port, join_host_port, make_zmq_path, make_async_with_semaphore, is_meta_module, is_non_overlapping_and_dense, supports_xpu_graph, create_kv_caches_with_random*, and the unused has_* capability probes. - Delete the unreachable deep-gemm mega-MoE warmup path, which imported vllm.models.deepseek_v4.nvidia.model and vllm.models.kimi_k3.nvidia.model, neither of which exists anymore. - Drop the imports and the two now-empty guards left behind. Verified statically: every changed file compiles, no unused import remains, and no deleted name is referenced in vllm, tests or csrc. The runtime gates still need the Python environment rebuilt. Co-authored-by: Hinata <hinata@idihore.id> Signed-off-by: MOVZX <movzx@yahoo.com>
222 lines
6.2 KiB
Python
222 lines
6.2 KiB
Python
# SPDX-License-Identifier: Apache-2.0
|
|
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
|
|
|
|
import os
|
|
from functools import wraps
|
|
from typing import TYPE_CHECKING
|
|
|
|
import torch
|
|
|
|
import vllm.envs as envs
|
|
from vllm.logger import init_logger
|
|
|
|
if TYPE_CHECKING:
|
|
pass
|
|
|
|
logger = init_logger(__name__)
|
|
|
|
try:
|
|
from amdsmi import (
|
|
amdsmi_get_gpu_asic_info,
|
|
amdsmi_get_processor_handles,
|
|
amdsmi_init,
|
|
amdsmi_shut_down,
|
|
)
|
|
except ImportError as e:
|
|
logger.warning("Failed to import from amdsmi with %r", e)
|
|
|
|
try:
|
|
import vllm._C # noqa: F401
|
|
except ImportError as e:
|
|
logger.warning("Failed to import from vllm._C with %r", e)
|
|
|
|
# import custom ops, trigger op registration
|
|
try:
|
|
import vllm._C_stable_libtorch # noqa: F401
|
|
except ImportError as e:
|
|
logger.warning("Failed to import from vllm._C_stable_libtorch with %r", e)
|
|
|
|
# Models not supported by ROCm.
|
|
_ROCM_UNSUPPORTED_MODELS: list[str] = []
|
|
|
|
# Models partially supported by ROCm.
|
|
# Architecture -> Reason.
|
|
_ROCM_PARTIALLY_SUPPORTED_MODELS: dict[str, str] = {}
|
|
_ROCM_DEVICE_ID_NAME_MAP: dict[str, str] = {
|
|
"0x74a0": "AMD_Instinct_MI300A",
|
|
"0x74a1": "AMD_Instinct_MI300X",
|
|
"0x74b5": "AMD_Instinct_MI300X", # MI300X VF
|
|
"0x74a2": "AMD_Instinct_MI308X",
|
|
"0x74a5": "AMD_Instinct_MI325X",
|
|
"0x74b9": "AMD_Instinct_MI325X", # MI325X VF
|
|
"0x74a9": "AMD_Instinct_MI300X_HF",
|
|
"0x74bd": "AMD_Instinct_MI300X_HF",
|
|
"0x744c": "AMD_Radeon_RX7900XTX",
|
|
# RDNA 3.5 APUs (Strix Point / Strix Halo)
|
|
"0x150e": "AMD_Radeon_890M", # gfx1150, Strix Point
|
|
"0x1586": "AMD_Radeon_8060S", # gfx1151, Strix Halo
|
|
# RDNA 4 discrete (Navi 48)
|
|
"0x7550": "AMD_Radeon_RX9070XT", # gfx1201
|
|
"0x7551": "AMD_Radeon_R9700", # gfx1201
|
|
}
|
|
|
|
|
|
def _sync_hip_cuda_env_vars():
|
|
"""Ensure HIP_VISIBLE_DEVICES and CUDA_VISIBLE_DEVICES are consistent.
|
|
Treats empty string as unset. Raises on genuine conflicts."""
|
|
hip_val = os.environ.get("HIP_VISIBLE_DEVICES") or None
|
|
cuda_val = os.environ.get("CUDA_VISIBLE_DEVICES") or None
|
|
|
|
if hip_val is not None and cuda_val is not None:
|
|
if hip_val != cuda_val:
|
|
raise ValueError(
|
|
f"Inconsistent GPU visibility env vars: "
|
|
f"HIP_VISIBLE_DEVICES='{hip_val}' vs "
|
|
f"CUDA_VISIBLE_DEVICES='{cuda_val}'. "
|
|
f"Please set only one, or ensure they match."
|
|
)
|
|
elif hip_val is not None:
|
|
os.environ["CUDA_VISIBLE_DEVICES"] = hip_val
|
|
|
|
|
|
# Sync at import time - catches misconfigurations from process start.
|
|
_sync_hip_cuda_env_vars()
|
|
|
|
|
|
# AMDSMI utils
|
|
# Note that NVML is not affected by `{CUDA/HIP}_VISIBLE_DEVICES`,
|
|
# all the related functions work on real physical device ids.
|
|
# the major benefit of using AMDSMI is that it will not initialize CUDA
|
|
|
|
|
|
def with_amdsmi_context(fn):
|
|
@wraps(fn)
|
|
def wrapper(*args, **kwargs):
|
|
amdsmi_init()
|
|
try:
|
|
return fn(*args, **kwargs)
|
|
finally:
|
|
amdsmi_shut_down()
|
|
|
|
return wrapper
|
|
|
|
|
|
@with_amdsmi_context
|
|
def _query_gcn_arch_from_amdsmi() -> str:
|
|
"""Query GCN arch from amdsmi. Raises if not available."""
|
|
handles = amdsmi_get_processor_handles()
|
|
if handles:
|
|
asic_info = amdsmi_get_gpu_asic_info(handles[0])
|
|
# Use target_graphics_version which contains the gfx name
|
|
# e.g., 'gfx942' for MI300X/MI325X
|
|
target_gfx = asic_info.get("target_graphics_version", "")
|
|
if target_gfx:
|
|
return target_gfx
|
|
raise RuntimeError("amdsmi did not return valid GCN arch")
|
|
|
|
|
|
def _get_gcn_arch() -> str:
|
|
"""Get GCN arch via amdsmi (no CUDA init), fallback to torch.cuda.
|
|
Called once at module level; result stored in _GCN_ARCH.
|
|
"""
|
|
try:
|
|
return _query_gcn_arch_from_amdsmi()
|
|
except Exception as e:
|
|
logger.debug("Failed to get GCN arch via amdsmi: %s", e)
|
|
# Ultimate fallback: use torch.cuda (will initialize CUDA)
|
|
return torch.cuda.get_device_properties("cuda").gcnArchName
|
|
|
|
|
|
# Resolve once at module load. Uses amdsmi (no CUDA init) so Ray workers
|
|
# can still set CUDA_VISIBLE_DEVICES after import.
|
|
# These are plain Python bools — fully torch.compile/Dynamo safe.
|
|
_GCN_ARCH = _get_gcn_arch()
|
|
|
|
_ON_GFX1X = any(arch in _GCN_ARCH for arch in ["gfx11", "gfx12"])
|
|
_ON_GFX11 = "gfx11" in _GCN_ARCH
|
|
_ON_GFX1100 = "gfx1100" in _GCN_ARCH
|
|
_ON_GFX1151 = "gfx1151" in _GCN_ARCH
|
|
_ON_GFX115X = "gfx115" in _GCN_ARCH
|
|
_ON_GFX12X = any(arch in _GCN_ARCH for arch in ["gfx12"])
|
|
_ON_MI3XX = any(arch in _GCN_ARCH for arch in ["gfx942", "gfx950"])
|
|
_ON_GFX9 = any(arch in _GCN_ARCH for arch in ["gfx90a", "gfx942", "gfx950"])
|
|
_ON_GFX90A = "gfx90a" in _GCN_ARCH
|
|
_ON_GFX942 = "gfx942" in _GCN_ARCH
|
|
_ON_GFX950 = "gfx950" in _GCN_ARCH
|
|
_ON_GFX1250 = "gfx1250" in _GCN_ARCH
|
|
|
|
_ON_CDNA = any(arch in _GCN_ARCH for arch in ["gfx9", "gfx1250"])
|
|
# RDNA = gfx11/gfx12 minus the CDNA-classified gfx1250.
|
|
_ON_RDNA = _ON_GFX1X and not _ON_CDNA
|
|
_ON_RDNA4 = any(arch in _GCN_ARCH for arch in ["gfx1200", "gfx1201"])
|
|
|
|
|
|
def on_gfx1x() -> bool:
|
|
return _ON_GFX1X and not _ON_CDNA
|
|
|
|
|
|
def on_gfx11() -> bool:
|
|
return _ON_GFX11
|
|
|
|
|
|
def on_gfx1100() -> bool:
|
|
return _ON_GFX1100
|
|
|
|
|
|
def on_gfx12x() -> bool:
|
|
return _ON_GFX12X and not _ON_CDNA
|
|
|
|
|
|
def on_gfx1250() -> bool:
|
|
return _ON_GFX1250
|
|
|
|
|
|
def on_rdna4() -> bool:
|
|
return _ON_RDNA4
|
|
|
|
|
|
def on_mi3xx() -> bool:
|
|
return _ON_MI3XX
|
|
|
|
|
|
def on_gfx90a() -> bool:
|
|
return _ON_GFX90A
|
|
|
|
|
|
def on_gfx942() -> bool:
|
|
return _ON_GFX942
|
|
|
|
|
|
def on_gfx950() -> bool:
|
|
return _ON_GFX950
|
|
|
|
|
|
def get_cdna_version() -> int:
|
|
if on_gfx90a():
|
|
return 2
|
|
if on_gfx942():
|
|
return 3
|
|
if on_gfx950():
|
|
return 4
|
|
if on_gfx1250():
|
|
return 5
|
|
return 0
|
|
|
|
|
|
# Enable HIP online tuning early, before hipBLASLt initializes.
|
|
# Turn on hipBLASLt online tuning if use AITER hipBLASLt GEMM.
|
|
if (
|
|
envs.VLLM_ROCM_USE_AITER
|
|
and envs.VLLM_ROCM_USE_AITER_LINEAR
|
|
and envs.VLLM_ROCM_USE_AITER_LINEAR_HIPBMM
|
|
and get_cdna_version() > 2
|
|
):
|
|
os.environ["HIP_ONLINE_TUNING"] = "1"
|
|
|
|
# RCCL 2.30's RAS thread can abort the process at exit ("terminate called
|
|
# without an active exception") when its shutdown notification fails.
|
|
# https://github.com/ROCm/rocm-systems/issues/13572
|
|
if torch.version.hip:
|
|
os.environ.setdefault("NCCL_RAS_ENABLE", "0")
|
|
|