rocmroll/profiles/sage-attention.json at 600760e2f3467eedab2e1327dc58c0a3b75ac586 · tomsnunes/rocmroll · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
{
  "name": "sage-attention",
  "description": "Full ROCm performance profile. Enables Sage Attention Triton backend, MIOpen kernel search, SageAttention, and dynamic VRAM. Disables smart memory and pinned memory for better AMD GPU performance. Default for the nightly channel.",
  "version": "1.0",
  "defaultForChannels": ["nightly"],
  "env": {
    "COMFYUI_ENABLE_MIOPEN": "0",
    "FLASH_ATTENTION_TRITON_AMD_ENABLE": "TRUE",
    "FLASH_ATTENTION_TRITON_AMD_AUTOTUNE": "TRUE",
    "MIOPEN_FIND_ENFORCE": "1",
    "MIOPEN_FIND_MODE": "2",
    "MIOPEN_DEBUG_DISABLE_FIND_DB": "0",
    "MIOPEN_SEARCH_CUTOFF": "1",
    "MIOPEN_ENABLE_LOGGING": "0",
    "MIOPEN_LOG_LEVEL": "0",
    "MIOPEN_ENABLE_LOGGING_CMD": "0",
    "TRITON_PRINT_AUTOTUNING": "0",
    "TRITON_CACHE_AUTOTUNING": "0",
    "PYTHONUNBUFFERED": "1",
    "PYTHONIOENCODING": "utf-8",
    "PYTHONUTF8": "1"
  },
  "launchArgs": [
    "--disable-api-nodes",
    "--disable-smart-memory",
    "--disable-pinned-memory",
    "--preview-method", "auto",
    "--use-sage-attention",
    "--enable-manager",
    "--enable-manager-legacy-ui",
    "--enable-dynamic-vram"
  ],
  "legacyGpuOverrides": {
    "env": {
      "TORCH_BACKENDS_CUDA_FLASH_SDP_ENABLED": "0",
      "TORCH_BACKENDS_CUDA_MEM_EFF_SDP_ENABLED": "0",
      "TORCH_BACKENDS_CUDA_MATH_SDP_ENABLED": "1"
    },
    "launchArgs": [
      "--use-quad-cross-attention"
    ]
  }
}