mirror of
https://github.com/FlashML-org/FreeToken.git
synced 2026-11-23 02:00:02 +08:00
* feat: add simple ROCm GPU support for RDNA3 (gfx1100-1103) - Add hip_compat.h shim mapping CUDA runtime API to HIP equivalents - Update pinned_tensor.cpp to compile under both nvcc and hipcc - Add ROCm detection in arch.py (is_rocm, get_rocm_gfx_arch, is_gfx11xx_family) - Guard NVIDIA arch checks to return None on ROCm - Skip nvcc version check in _toolchain.py when on ROCm - Add ROCm build path in setup.py (ROCM_HOME, amdhip64, --offload-arch) - Add _hip_cflags() in kernel/utils.py for JIT compilation on ROCm - Add is_rocm() and driver_hip_version() in backend.py - Add rocm-smi fallback in __main__.py for clangd generation - Add TODO(ROCm) for NCCL->RCCL, flashinfer/sgl_kernel ROCm builds, Triton autotune RDNA3 tuning, PDL equivalent, hiprtc JIT cache - Add AMD ROCm classifier in pyproject.toml * fix(rocm): fall back for single-bank fast index copy * test(rocm): cover fast index copy fallback * test(rocm): use explicit fast copy module import * fix(rocm): complete RDNA3 runtime path * fix(rocm): preserve CPU MoE implementation * build(rocm): support RDNA4 discovery and modular SDKs * fix(rocm): complete modular SDK runtime support * test(rocm): harden foundation platform contracts * fix(rocm): pin jit target and split install guide * perf(kernel): restore ROCm non-temporal index copy --------- Co-authored-by: bouclem <bouclement2012@gmail.com> Co-authored-by: Yuu <206304251+nekomario28@users.noreply.github.com> Co-authored-by: LZ-QWQ <lz1830837889@163.com>
83 lines
2.6 KiB
Python
83 lines
2.6 KiB
Python
"""Availability probes for the optional native kernel packages.
|
|
|
|
When flashinfer / sgl_kernel are installed the call-sites use their fused CUDA
|
|
ops; otherwise they fall back to the pure-Triton kernels in
|
|
``freetoken.kernel.triton``. ``find_spec`` only checks that the package is
|
|
importable (no import side effects), and the result is cached.
|
|
"""
|
|
from __future__ import annotations
|
|
|
|
import functools
|
|
import importlib.util
|
|
|
|
import torch
|
|
|
|
|
|
def _importable(name: str) -> bool:
|
|
# find_spec normally returns None when a package is absent, but it can raise
|
|
# (broken parent package, or a meta_path finder that blocks the name); treat
|
|
# any failure as "not available" so callers cleanly fall back to triton.
|
|
try:
|
|
return importlib.util.find_spec(name) is not None
|
|
except Exception:
|
|
return False
|
|
|
|
|
|
@functools.cache
|
|
def is_flashinfer_installed() -> bool:
|
|
return _importable("flashinfer")
|
|
|
|
|
|
@functools.cache
|
|
def is_sgl_kernel_installed() -> bool:
|
|
return _importable("sgl_kernel")
|
|
|
|
|
|
@functools.cache
|
|
def is_vllm_installed() -> bool:
|
|
return _importable("vllm")
|
|
|
|
|
|
@functools.cache
|
|
def device_capability() -> tuple[int, int]:
|
|
"""Compute capability of the current device as (major, minor); (0, 0) without CUDA."""
|
|
if not torch.cuda.is_available():
|
|
return (0, 0)
|
|
major, minor = torch.cuda.get_device_capability()
|
|
return (int(major), int(minor))
|
|
|
|
|
|
@functools.cache
|
|
def is_rocm() -> bool:
|
|
"""True when torch is built for ROCm (AMD GPU)."""
|
|
import torch
|
|
return getattr(torch.version, "hip", None) is not None
|
|
|
|
|
|
@functools.cache
|
|
def driver_hip_version() -> int | None:
|
|
"""ROCm driver version, or None if undetermined."""
|
|
# TODO(ROCm): flashinfer/sgl_kernel have no ROCm builds — Triton fallback is used.
|
|
try:
|
|
from freetoken.kernel.pinned import _load_pinned_extension
|
|
return int(_load_pinned_extension().driver_cuda_version()) or None
|
|
except Exception:
|
|
return None
|
|
|
|
|
|
@functools.cache
|
|
def driver_cuda_version() -> int | None:
|
|
"""Max CUDA version the installed NVIDIA driver supports (``13000`` == CUDA 13.0),
|
|
or None if undetermined. Driver-JIT kernels (PTX compiled at runtime, e.g.
|
|
flashinfer's CuTe-DSL paths) are gated by this, not by any package's build-time
|
|
toolkit version. Resolved through the ``_pinned_tensor`` extension's link-time
|
|
cudart, so it works wherever the extension builds (including Windows) -- no dlopen
|
|
by soname."""
|
|
try:
|
|
from freetoken.kernel.pinned import _load_pinned_extension
|
|
|
|
version = int(_load_pinned_extension().driver_cuda_version())
|
|
except Exception:
|
|
return None
|
|
return version or None # 0 == no driver installed
|