| 1 |
"""Hardware-capability mocks for doctor and planner tests. |
| 2 |
|
| 3 |
Each context manager flips a consistent set of `torch.*` attributes so |
| 4 |
code paths gated on `torch.cuda.is_available()`, `get_device_capability()`, |
| 5 |
`mem_get_info()`, MPS availability, or `torch.version.hip` behave as if |
| 6 |
the requested backend is present — without real hardware. |
| 7 |
|
| 8 |
`torch` is imported inside each function so merely collecting the module |
| 9 |
never touches torch state. |
| 10 |
""" |
| 11 |
|
| 12 |
from __future__ import annotations |
| 13 |
|
| 14 |
import contextlib |
| 15 |
from collections.abc import Iterator |
| 16 |
from unittest.mock import patch |
| 17 |
|
| 18 |
|
| 19 |
@contextlib.contextmanager |
| 20 |
def force_cuda( |
| 21 |
sm: tuple[int, int] = (8, 0), |
| 22 |
vram_gb: float = 24.0, |
| 23 |
device_name: str = "NVIDIA GeForce RTX 4090", |
| 24 |
) -> Iterator[None]: |
| 25 |
"""Pretend a CUDA GPU with `sm` compute capability and `vram_gb` free. |
| 26 |
|
| 27 |
`torch.cuda.mem_get_info()` returns (free, total) in bytes; we report |
| 28 |
the same value for both to make arithmetic simple. |
| 29 |
""" |
| 30 |
import torch |
| 31 |
|
| 32 |
free_bytes = int(vram_gb * (1024**3)) |
| 33 |
total_bytes = free_bytes |
| 34 |
patches = [ |
| 35 |
patch.object(torch.cuda, "is_available", return_value=True), |
| 36 |
patch.object(torch.cuda, "device_count", return_value=1), |
| 37 |
patch.object(torch.cuda, "get_device_name", return_value=device_name), |
| 38 |
patch.object(torch.cuda, "get_device_capability", return_value=sm), |
| 39 |
patch.object(torch.cuda, "mem_get_info", return_value=(free_bytes, total_bytes)), |
| 40 |
# ROCm build attribute must be absent on a "real" CUDA box. |
| 41 |
patch.object(torch.version, "hip", None), |
| 42 |
patch.object(torch.backends.mps, "is_available", return_value=False), |
| 43 |
patch.object(torch.backends.mps, "is_built", return_value=False), |
| 44 |
] |
| 45 |
with contextlib.ExitStack() as stack: |
| 46 |
for p in patches: |
| 47 |
stack.enter_context(p) |
| 48 |
yield |
| 49 |
|
| 50 |
|
| 51 |
@contextlib.contextmanager |
| 52 |
def force_rocm( |
| 53 |
vram_gb: float = 16.0, |
| 54 |
device_name: str = "AMD Radeon RX 7900 XTX", |
| 55 |
hip_version: str = "6.0", |
| 56 |
sm: tuple[int, int] = (11, 0), # HIP compute capability (RDNA3 ≈ 11.0.3) |
| 57 |
gcn_arch_name: str = "gfx1100", |
| 58 |
) -> Iterator[None]: |
| 59 |
"""Pretend a ROCm GPU. `torch.version.hip` is the distinguishing mark. |
| 60 |
|
| 61 |
`gcn_arch_name` (Sprint 22) is the AMD arch string — `gfx90a` |
| 62 |
(MI200), `gfx942` (MI300), `gfx1100` (RDNA3), `gfx1030` (RDNA2), |
| 63 |
etc. The bf16 + FlashAttention probes allowlist against this |
| 64 |
string rather than the CUDA-style SM tuple. |
| 65 |
""" |
| 66 |
from types import SimpleNamespace |
| 67 |
|
| 68 |
import torch |
| 69 |
|
| 70 |
free_bytes = int(vram_gb * (1024**3)) |
| 71 |
total_bytes = free_bytes |
| 72 |
device_props = SimpleNamespace( |
| 73 |
name=device_name, |
| 74 |
gcnArchName=gcn_arch_name, |
| 75 |
total_memory=total_bytes, |
| 76 |
) |
| 77 |
patches = [ |
| 78 |
patch.object(torch.cuda, "is_available", return_value=True), |
| 79 |
patch.object(torch.cuda, "device_count", return_value=1), |
| 80 |
patch.object(torch.cuda, "get_device_name", return_value=device_name), |
| 81 |
patch.object(torch.cuda, "get_device_capability", return_value=sm), |
| 82 |
patch.object(torch.cuda, "get_device_properties", return_value=device_props), |
| 83 |
patch.object(torch.cuda, "mem_get_info", return_value=(free_bytes, total_bytes)), |
| 84 |
patch.object(torch.version, "hip", hip_version), |
| 85 |
patch.object(torch.backends.mps, "is_available", return_value=False), |
| 86 |
patch.object(torch.backends.mps, "is_built", return_value=False), |
| 87 |
] |
| 88 |
with contextlib.ExitStack() as stack: |
| 89 |
for p in patches: |
| 90 |
stack.enter_context(p) |
| 91 |
yield |
| 92 |
|
| 93 |
|
| 94 |
@contextlib.contextmanager |
| 95 |
def force_mps() -> Iterator[None]: |
| 96 |
"""Pretend Apple Silicon (MPS backend available, no CUDA).""" |
| 97 |
import torch |
| 98 |
|
| 99 |
patches = [ |
| 100 |
patch.object(torch.cuda, "is_available", return_value=False), |
| 101 |
patch.object(torch.version, "hip", None), |
| 102 |
patch.object(torch.backends.mps, "is_available", return_value=True), |
| 103 |
patch.object(torch.backends.mps, "is_built", return_value=True), |
| 104 |
] |
| 105 |
with contextlib.ExitStack() as stack: |
| 106 |
for p in patches: |
| 107 |
stack.enter_context(p) |
| 108 |
yield |
| 109 |
|
| 110 |
|
| 111 |
@contextlib.contextmanager |
| 112 |
def force_cpu() -> Iterator[None]: |
| 113 |
"""Pretend CPU-only (no CUDA, no MPS).""" |
| 114 |
import torch |
| 115 |
|
| 116 |
patches = [ |
| 117 |
patch.object(torch.cuda, "is_available", return_value=False), |
| 118 |
patch.object(torch.version, "hip", None), |
| 119 |
patch.object(torch.backends.mps, "is_available", return_value=False), |
| 120 |
patch.object(torch.backends.mps, "is_built", return_value=False), |
| 121 |
] |
| 122 |
with contextlib.ExitStack() as stack: |
| 123 |
for p in patches: |
| 124 |
stack.enter_context(p) |
| 125 |
yield |