2780bc459a
The default Darwin path used to be a try/except ImportError, which only catches build failures. With the mtlgemm round 1 fixes shipped, the more common failure mode for end users will be running an *older* mtlgemm that still returns CPU tensors from MPS calls — that doesn't fail import, it fails with a cryptic LayerNorm crash inside the model on the first conv. The new probe runs a tiny SparseConv3d on MPS and checks the output device. If anything breaks (import, build, dispatch, return-device), fall back to the pure-PyTorch backend rather than crashing inside the model. Tensors in the probe are built on CPU then moved to MPS because some PyTorch builds lack int/fp16 torch.zeros kernels on MPS — that's a PyTorch issue, separate from anything we control here. Plus: end-to-end smoke test (test_flex_gemm_integration.py) that exercises both Algorithm.IMPLICIT_GEMM and Algorithm.MASKED_IMPLICIT_GEMM through a real SparseConv3d → F.layer_norm chain on MPS. Confirms both algorithms return MPS tensors of the right dtype and produce equivalent output. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
90 lines
2.8 KiB
Python
90 lines
2.8 KiB
Python
from typing import *
|
|
import platform
|
|
import sys
|
|
|
|
CONV = 'flex_gemm'
|
|
DEBUG = False
|
|
ATTN = 'flash_attn'
|
|
|
|
def __detect_defaults():
|
|
"""Auto-detect best backends for current platform."""
|
|
global CONV, ATTN
|
|
if platform.system() == 'Darwin':
|
|
ATTN = 'sdpa'
|
|
if __flex_gemm_works_on_mps():
|
|
CONV = 'flex_gemm'
|
|
else:
|
|
CONV = 'pytorch'
|
|
elif not __has_cuda():
|
|
CONV = 'pytorch'
|
|
ATTN = 'sdpa'
|
|
|
|
|
|
def __flex_gemm_works_on_mps():
|
|
"""Probe flex_gemm with a tiny MPS conv. If the install pre-dates the
|
|
device-routing fix (or fails to build), it returns a CPU tensor — fall
|
|
back to the pure-PyTorch backend rather than crashing inside the model
|
|
on the first LayerNorm. Build tensors on CPU and move to MPS because some
|
|
PyTorch builds lack int/fp16 torch.zeros kernels on MPS."""
|
|
try:
|
|
import torch
|
|
if not torch.backends.mps.is_available():
|
|
return False
|
|
import flex_gemm
|
|
from flex_gemm.ops.spconv import sparse_submanifold_conv3d, Algorithm, set_algorithm
|
|
set_algorithm(Algorithm.IMPLICIT_GEMM)
|
|
coords = torch.tensor([[0, 0, 0, 0]], dtype=torch.int32).to('mps')
|
|
feats = torch.zeros((1, 4), dtype=torch.float16).to('mps')
|
|
weight = torch.zeros((4, 1, 1, 1, 4), dtype=torch.float16).to('mps')
|
|
out, _ = sparse_submanifold_conv3d(feats, coords, torch.Size([1, 4, 1, 1, 1]), weight)
|
|
return out.device.type == 'mps'
|
|
except Exception:
|
|
return False
|
|
|
|
def __has_cuda():
|
|
try:
|
|
import torch
|
|
return torch.cuda.is_available()
|
|
except Exception:
|
|
return False
|
|
|
|
def __from_env():
|
|
import os
|
|
|
|
global CONV
|
|
global DEBUG
|
|
global ATTN
|
|
|
|
__detect_defaults()
|
|
|
|
env_sparse_conv_backend = os.environ.get('SPARSE_CONV_BACKEND')
|
|
env_sparse_debug = os.environ.get('SPARSE_DEBUG')
|
|
env_sparse_attn_backend = os.environ.get('SPARSE_ATTN_BACKEND')
|
|
if env_sparse_attn_backend is None:
|
|
env_sparse_attn_backend = os.environ.get('ATTN_BACKEND')
|
|
|
|
if env_sparse_conv_backend is not None and env_sparse_conv_backend in ['none', 'spconv', 'torchsparse', 'flex_gemm', 'pytorch']:
|
|
CONV = env_sparse_conv_backend
|
|
if env_sparse_debug is not None:
|
|
DEBUG = env_sparse_debug == '1'
|
|
if env_sparse_attn_backend is not None and env_sparse_attn_backend in ['xformers', 'flash_attn', 'flash_attn_3', 'sdpa']:
|
|
ATTN = env_sparse_attn_backend
|
|
|
|
print(f"[SPARSE] Conv backend: {CONV}; Attention backend: {ATTN}")
|
|
|
|
|
|
__from_env()
|
|
|
|
|
|
def set_conv_backend(backend: Literal['none', 'spconv', 'torchsparse', 'flex_gemm', 'pytorch']):
|
|
global CONV
|
|
CONV = backend
|
|
|
|
def set_debug(debug: bool):
|
|
global DEBUG
|
|
DEBUG = debug
|
|
|
|
def set_attn_backend(backend: Literal['xformers', 'flash_attn', 'sdpa']):
|
|
global ATTN
|
|
ATTN = backend
|