Public/automatic
1
0
Fork 0
mirror of https://github.com/vladmandic/automatic synced 2026-09-07 21:40:46 +02:00
Code Issues Packages Projects Releases Wiki Activity

Added further rocblas support enhancements and performance-related best practice settings.

Browse Source
This commit is contained in:
resonantsky
2026-04-02 14:59:00 +02:00
parent 668a94141d
commit f5c037a735
3 changed files with 221 additions and 27 deletions
Download Patch File Download Diff File Expand all files Collapse all files
scripts/rocm/rocm_mgr.py
+43 -1
View File
@@ -1,4 +1,5 @@
import os
import re
import sys
from pathlib import Path
from typing import Dict, Optional
@@ -121,14 +122,22 @@ def _get_venv() -> str:
return os.environ.get("VIRTUAL_ENV", "") or sys.prefix
def _get_root() -> str:
"""App root — one level above the venv folder (e.g. E:\\Sd.Next)."""
return str(Path(_get_venv()).parent)
def _expand_venv(value: str) -> str:
return value.replace("{VIRTUAL_ENV}", _get_venv())
return value.replace("{VIRTUAL_ENV}", _get_venv()).replace("{ROOT}", _get_root())
def _collapse_venv(value: str) -> str:
venv = _get_venv()
root = _get_root()
if venv and value.startswith(venv):
return "{VIRTUAL_ENV}" + value[len(venv):]
if root and value.startswith(root):
return "{ROOT}" + value[len(root):]
return value
@@ -365,6 +374,28 @@ def _user_db_summary(path: Path) -> dict:
return out
def _extract_db_hash(db_path: Path) -> str:
"""Derive the cache subfolder name from udb.txt filenames.
e.g. gfx1030_30.HIP.3_5_1_5454e9e2da.udb.txt → '3.5.1.5454e9e2da'"""
for f in db_path.glob("*.HIP.*.udb.txt"):
m = re.search(r'\.HIP\.([^.]+)\.udb\.txt$', f.name)
if m:
return m.group(1).replace("_", ".")
return ""
def _user_cache_summary(path: Path) -> dict:
"""Return {filename: 'N KB'} for binary cache blobs in the resolved cache path."""
out = {}
if not path.exists():
return out
for f in sorted(path.iterdir()):
if f.is_file():
kb = f.stat().st_size // 1024
out[f.name] = f"{kb} KB"
return out
def info() -> dict:
config = load_config()
db_path = Path(_expand_venv(config.get("MIOPEN_SYSTEM_DB_PATH", "")))
@@ -427,12 +458,23 @@ def info() -> dict:
if ufiles:
udb["files"] = ufiles
# --- User cache (~/.miopen/cache/<version-hash>) ---
cache_base = Path.home() / ".miopen" / "cache"
db_hash = _extract_db_hash(user_db_path) if user_db_path.exists() else ""
cache_path = cache_base / db_hash if db_hash else cache_base
ucache = {"path": str(cache_path), "exists": cache_path.exists()}
if cache_path.exists():
cfiles = _user_cache_summary(cache_path)
if cfiles:
ucache["files"] = cfiles
return {
"rocm": rocm_section,
"torch": torch_section,
"gpu": gpu_section,
"system_db": sdb,
"user_db": udb,
"user_cache": ucache,
}
scripts/rocm/rocm_vars.py
+112 -7
View File
@@ -3,6 +3,7 @@ from typing import Dict, Any, List, Tuple
# --- General MIOpen/rocBLAS variables (dropdown/textbox/checkbox) ---
GENERAL_VARS: Dict[str, Dict[str, Any]] = {
# ── GEMM backend selector + companion toggles ──────────────────────────
"MIOPEN_GEMM_ENFORCE_BACKEND": {
"default": "1",
"desc": "Enforce GEMM backend",
@@ -10,6 +11,29 @@ GENERAL_VARS: Dict[str, Dict[str, Any]] = {
"options": [("1 - rocBLAS", "1"), ("5 - hipBLASLt", "5")],
"restart_required": False,
},
"PYTORCH_ROCM_USE_ROCBLAS": {
"default": "1",
"desc": "PyTorch ROCm: prioritise rocBLAS for linear algebra",
"widget": "dropdown",
"options": [("0 - Off", "0"), ("1 - On", "1")],
"restart_required": True,
},
"PYTORCH_HIPBLASLT_DISABLE": {
"default": "1",
"desc": "Disable PyTorch hipBLASLt dispatcher",
"widget": "dropdown",
"options": [("0 - Allow hipBLASLt", "0"), ("1 - Disable hipBLASLt", "1")],
"restart_required": True,
},
"ROCBLAS_USE_HIPBLASLT": {
"default": "0",
"desc": "rocBLAS: use hipBLASLt backend (0 = Tensile)",
"widget": "dropdown",
"options": [("0 - Tensile (rocBLAS)", "0"), ("1 - hipBLASLt", "1")],
"restart_required": True,
},
# ── MIOpen behavioural settings ────────────────────────────────────────
"MIOPEN_FIND_MODE": {
"default": "2",
"desc": "MIOpen Find Mode",
@@ -31,6 +55,15 @@ GENERAL_VARS: Dict[str, Dict[str, Any]] = {
"options": [("0 - Off", "0"), ("1 - On", "1")],
"restart_required": True,
},
"MIOPEN_DEBUG_CONVOLUTION_DETERMINISTIC": {
"default": "0",
"desc": "Deterministic convolution (reproducible results, may be slower)",
"widget": "dropdown",
"options": [("0 - Off", "0"), ("1 - On", "1")],
"restart_required": False,
},
# ── Paths / sizes ──────────────────────────────────────────────────────
"MIOPEN_SYSTEM_DB_PATH": {
"default": "{VIRTUAL_ENV}\\Lib\\site-packages\\_rocm_sdk_devel\\bin\\",
"desc": "MIOpen system DB path",
@@ -38,6 +71,75 @@ GENERAL_VARS: Dict[str, Dict[str, Any]] = {
"options": None,
"restart_required": True,
},
"MIOPEN_CONVOLUTION_MAX_WORKSPACE": {
"default": "1073741824",
"desc": "MIOpen convolution max workspace (bytes; 1 GB default)",
"widget": "textbox",
"options": None,
"restart_required": False,
},
"ROCBLAS_TENSILE_LIBPATH": {
"default": "{VIRTUAL_ENV}\\Lib\\site-packages\\_rocm_sdk_devel\\bin\\rocblas\\library",
"desc": "rocBLAS Tensile library path",
"widget": "textbox",
"options": None,
"restart_required": True,
},
"ROCBLAS_DEVICE_MEMORY_SIZE": {
"default": "",
"desc": "rocBLAS workspace size in bytes (empty = dynamic)",
"widget": "textbox",
"options": None,
"restart_required": False,
},
"PYTORCH_TUNABLEOP_CACHE_DIR": {
"default": "{ROOT}\\models\\tunable",
"desc": "TunableOp: kernel profile cache directory",
"widget": "textbox",
"options": None,
"restart_required": False,
},
# ── rocBLAS settings ───────────────────────────────────────────────────
"ROCBLAS_STREAM_ORDER_ALLOC": {
"default": "1",
"desc": "rocBLAS stream-ordered memory allocation",
"widget": "dropdown",
"options": [("0 - Standard", "0"), ("1 - Stream-ordered", "1")],
"restart_required": False,
},
"ROCBLAS_DEFAULT_ATOMICS_MODE": {
"default": "1",
"desc": "rocBLAS default atomics mode (1 = allow non-deterministic for performance)",
"widget": "dropdown",
"options": [("0 - Off (deterministic)", "0"), ("1 - On (performance)", "1")],
"restart_required": False,
},
"PYTORCH_TUNABLEOP_ROCBLAS_ENABLED": {
"default": "1",
"desc": "TunableOp: wrap and optimise rocBLAS GEMM calls",
"widget": "dropdown",
"options": [("0 - Off", "0"), ("1 - On", "1")],
"restart_required": False,
},
"PYTORCH_TUNABLEOP_TUNING": {
"default": "0",
"desc": "TunableOp: tuning mode (1 = benchmark; 0 = use saved CSV)",
"widget": "dropdown",
"options": [("0 - Use saved CSV", "0"), ("1 - Benchmark new shapes", "1")],
"restart_required": False,
},
# ── hipBLASLt settings ─────────────────────────────────────────────────
"PYTORCH_TUNABLEOP_HIPBLASLT_ENABLED": {
"default": "0",
"desc": "TunableOp: benchmark hipBLASLt kernels",
"widget": "dropdown",
"options": [("0 - Off", "0"), ("1 - On", "1")],
"restart_required": False,
},
# ── Logging: MIOpen → rocBLAS → hipBLASLt ─────────────────────────────
"MIOPEN_LOG_LEVEL": {
"default": "0",
"desc": "MIOpen log verbosity level",
@@ -66,13 +168,6 @@ GENERAL_VARS: Dict[str, Dict[str, Any]] = {
"options": [("0 - Off", "0"), ("1 - Error", "1"), ("2 - Trace", "2"), ("3 - Hints", "3"), ("4 - Info", "4"), ("5 - API Trace", "5")],
"restart_required": False,
},
"MIOPEN_DEBUG_CONVOLUTION_DETERMINISTIC": {
"default": "0",
"desc": "Deterministic convolution (reproducible results, may be slower)",
"widget": "dropdown",
"options": [("0 - Off", "0"), ("1 - On", "1")],
"restart_required": False,
},
}
# --- Solver toggles (inference/FWD only, RDNA2/3/4 compatible) ---
@@ -251,3 +346,13 @@ SOLVER_GROUPS: List[Tuple[str, List[str]]] = [
"MIOPEN_DEBUG_CK_DEFAULT_KERNELS",
]),
]
# Variables that are relevant only when hipBLASLt is the active GEMM backend.
# These are visually greyed-out in the UI when rocBLAS (MIOPEN_GEMM_ENFORCE_BACKEND="1") is selected.
HIPBLASLT_VARS: set = {
"PYTORCH_HIPBLASLT_DISABLE",
"ROCBLAS_USE_HIPBLASLT",
"PYTORCH_TUNABLEOP_HIPBLASLT_ENABLED",
"HIPBLASLT_LOG_LEVEL",
}