mirror of
https://github.com/comfyanonymous/ComfyUI.git
synced 2026-04-15 21:12:30 +08:00
entrypoint: derive correct arch list; add user-tunable build parallelism; fix Sage flags; first-run installs
- Auto-derive TORCH_CUDA_ARCH_LIST from torch device capabilities (unique, sorted, optional +PTX) to cover all charted GPUs: Turing 7.5, Ampere 8.0/8.6/8.7, Ada 8.9, Hopper 9.0, and Blackwell 10.0 & 12.0/12.1; add name-based fallbacks for mixed or torch-less scenarios. - Add user-tunable build parallelism with SAGE_MAX_JOBS (preferred) and MAX_JOBS (alias) that cap PyTorch cpp_extension/ninja -j; fall back to a RAM/CPU heuristic to prevent OOM “Killed” during CUDA/C++ builds. - Correct Sage flags: SAGE_ATTENTION_AVAILABLE only signals “built/installed,” while FORCE_SAGE_ATTENTION=1 enables Sage at startup; fix logs to reference FORCE_SAGE_ATTENTION. - Maintain Triton install strategy by GPU generation for compatibility and performance. - Add first-run dependency installation with COMFY_FORCE_INSTALL override; keep permissions bootstrap and minor logging/URL cleanups.
This commit is contained in:
clsferguson
GitHub
parent
45b87c7c99
commit
f6d49f33b7
435
entrypoint.sh
435
entrypoint.sh
@ -1,6 +1,7 @@
|
||||
#!/bin/bash
|
||||
set -euo pipefail
|
||||
|
||||
# --- config ---
|
||||
APP_USER=${APP_USER:-appuser}
|
||||
APP_GROUP=${APP_GROUP:-appuser}
|
||||
PUID=${PUID:-1000}
|
||||
@ -12,379 +13,311 @@ SAGE_ATTENTION_BUILT_FLAG="$SAGE_ATTENTION_DIR/.built"
|
||||
PERMISSIONS_SET_FLAG="$BASE_DIR/.permissions_set"
|
||||
FIRST_RUN_FLAG="$BASE_DIR/.first_run_done"
|
||||
|
||||
# Function to log with timestamp
|
||||
log() {
|
||||
echo "[$(date '+%H:%M:%S')] $1"
|
||||
# Sage flags:
|
||||
# - SAGE_ATTENTION_AVAILABLE: set by script to indicate built/importable
|
||||
# - FORCE_SAGE_ATTENTION=1: force app to enable at startup (default 0)
|
||||
|
||||
# --- logging ---
|
||||
log() { echo "[$(date '+%H:%M:%S')] $1"; }
|
||||
|
||||
# --- build parallelism (single knob) ---
|
||||
# Public knob: SAGE_MAX_JOBS. If unset, pick RAM/CPU heuristic.
|
||||
decide_build_jobs() {
|
||||
if [ -n "${SAGE_MAX_JOBS:-}" ]; then echo "$SAGE_MAX_JOBS"; return; fi
|
||||
local mem_kb=$(awk '/MemTotal:/ {print $2}' /proc/meminfo 2>/dev/null || echo 0)
|
||||
local cpu=$(nproc) cap=24 jobs
|
||||
if [ "$mem_kb" -le $((8*1024*1024)) ]; then jobs=2
|
||||
elif [ "$mem_kb" -le $((12*1024*1024)) ]; then jobs=3
|
||||
elif [ "$mem_kb" -le $((24*1024*1024)) ]; then jobs=4
|
||||
elif [ "$mem_kb" -le $((64*1024*1024)) ]; then jobs=$(( cpu<8 ? cpu : 8 ))
|
||||
else jobs=$cpu; [ "$jobs" -gt "$cap" ] && jobs=$cap
|
||||
fi
|
||||
echo "$jobs"
|
||||
}
|
||||
|
||||
# Function to test PyTorch CUDA compatibility
|
||||
# --- CUDA/Torch checks ---
|
||||
test_pytorch_cuda() {
|
||||
python -c "
|
||||
import torch, sys
|
||||
if not torch.cuda.is_available():
|
||||
print('[ERROR] PyTorch CUDA not available')
|
||||
sys.exit(1)
|
||||
c = torch.cuda.device_count()
|
||||
print(f'[TEST] PyTorch CUDA available with {c} devices')
|
||||
print('[ERROR] PyTorch CUDA not available'); sys.exit(1)
|
||||
c=torch.cuda.device_count(); print(f'[TEST] PyTorch CUDA available with {c} devices')
|
||||
for i in range(c):
|
||||
props = torch.cuda.get_device_properties(i)
|
||||
print(f'[TEST] GPU {i}: {props.name} (Compute {props.major}.{props.minor})')
|
||||
p=torch.cuda.get_device_properties(i)
|
||||
print(f'[TEST] GPU {i}: {p.name} (Compute {p.major}.{p.minor})')
|
||||
" 2>/dev/null
|
||||
}
|
||||
|
||||
# Function to detect all GPUs and their generations
|
||||
# Derive arch list directly from Torch; optional +PTX via SAGE_PTX_FALLBACK=1
|
||||
compute_arch_list_from_torch() {
|
||||
python - <<'PY' 2>/dev/null
|
||||
import os, sys
|
||||
try:
|
||||
import torch
|
||||
if not torch.cuda.is_available():
|
||||
print(""); sys.exit(0)
|
||||
caps = {f"{torch.cuda.get_device_properties(i).major}.{torch.cuda.get_device_properties(i).minor}"
|
||||
for i in range(torch.cuda.device_count())}
|
||||
ordered = sorted(caps, key=lambda s: tuple(int(x) for x in s.split(".")))
|
||||
if not ordered: print(""); sys.exit(0)
|
||||
if os.environ.get("SAGE_PTX_FALLBACK","0")=="1":
|
||||
highest = ordered[-1]; print(";".join(ordered+[highest + "+PTX"]))
|
||||
else:
|
||||
print(";".join(ordered))
|
||||
except Exception:
|
||||
print("")
|
||||
PY
|
||||
}
|
||||
|
||||
# Fallback name-based mapping across Turing→Blackwell
|
||||
detect_gpu_generations() {
|
||||
local gpu_info
|
||||
gpu_info=$(nvidia-smi --query-gpu=name --format=csv,noheader,nounits 2>/dev/null || echo "")
|
||||
local has_rtx20=false
|
||||
local has_rtx30=false
|
||||
local has_rtx40=false
|
||||
local has_rtx50=false
|
||||
local gpu_count=0
|
||||
|
||||
if [ -z "$gpu_info" ]; then
|
||||
log "No NVIDIA GPUs detected"
|
||||
return 1
|
||||
fi
|
||||
|
||||
local info=$(nvidia-smi --query-gpu=name --format=csv,noheader,nounits 2>/dev/null || echo "")
|
||||
local has_turing=false has_amp_ga100=false has_amp_ga10x=false has_amp_ga10b=false
|
||||
local has_ada=false has_hopper=false has_bw_cons=false has_bw_dc=false
|
||||
local n=0
|
||||
[ -z "$info" ] && { log "No NVIDIA GPUs detected"; return 1; }
|
||||
log "Detecting GPU generations:"
|
||||
while IFS= read -r gpu; do
|
||||
gpu_count=$((gpu_count + 1))
|
||||
log " GPU $gpu_count: $gpu"
|
||||
case "$gpu" in
|
||||
*"RTX 20"*|*"2060"*|*"2070"*|*"2080"*|*"2090"*) has_rtx20=true ;;
|
||||
*"RTX 30"*|*"3060"*|*"3070"*|*"3080"*|*"3090"*) has_rtx30=true ;;
|
||||
*"RTX 40"*|*"4060"*|*"4070"*|*"4080"*|*"4090"*) has_rtx40=true ;;
|
||||
*"RTX 50"*|*"5060"*|*"5070"*|*"5080"*|*"5090"*) has_rtx50=true ;;
|
||||
while IFS= read -r g; do
|
||||
n=$((n+1)); log " GPU $n: $g"
|
||||
case "$g" in
|
||||
*"RTX 20"*|*"T4"*) has_turing=true ;;
|
||||
*"A100"*|*"A30"*|*"A40"*) has_amp_ga100=true ;;
|
||||
*"RTX 30"*|*"RTX 3090"*|*"RTX 3080"*|*"RTX 3070"*|*"RTX 3060"*) has_amp_ga10x=true ;;
|
||||
*"Orin"*|*"Jetson"*) has_amp_ga10b=true ;;
|
||||
*"RTX 40"*|*"4090"*|*"L40"*|*"L4"*) has_ada=true ;;
|
||||
*"H100"*|*"H200"*|*"GH200"*) has_hopper=true ;;
|
||||
*"RTX 50"*|*"5090"*|*"5080"*|*"5070"*|*"5060"*|*"PRO "*Blackwell*|*"PRO 4000 Blackwell"*) has_bw_cons=true ;;
|
||||
*"B200"*|*"B100"*|*"GB200"*|*"B40"*|*"RTX 6000 Blackwell"*|*"RTX 5000 Blackwell"*) has_bw_dc=true ;;
|
||||
esac
|
||||
done <<< "$gpu_info"
|
||||
|
||||
export DETECTED_RTX20=$has_rtx20
|
||||
export DETECTED_RTX30=$has_rtx30
|
||||
export DETECTED_RTX40=$has_rtx40
|
||||
export DETECTED_RTX50=$has_rtx50
|
||||
export GPU_COUNT=$gpu_count
|
||||
|
||||
log "Detection summary: RTX20=$has_rtx20, RTX30=$has_rtx30, RTX40=$has_rtx40, RTX50=$has_rtx50"
|
||||
|
||||
if test_pytorch_cuda; then
|
||||
log "PyTorch CUDA compatibility confirmed"
|
||||
else
|
||||
log "WARNING: PyTorch CUDA compatibility issues detected"
|
||||
fi
|
||||
done <<< "$info"
|
||||
export DET_TURING=$has_turing DET_AMP80=$has_amp_ga100 DET_AMP86=$has_amp_ga10x DET_AMP87=$has_amp_ga10b
|
||||
export DET_ADA=$has_ada DET_HOPPER=$has_hopper DET_BW12=$has_bw_cons DET_BW10=$has_bw_dc
|
||||
export GPU_COUNT=$n
|
||||
log "Summary: Turing=$has_turing Amp(8.0)=$has_amp_ga100 Amp(8.6)=$has_amp_ga10x Amp(8.7)=$has_amp_ga10b Ada=$has_ada Hopper=$has_hopper Blackwell(12.x)=$has_bw_cons Blackwell(10.0)=$has_bw_dc"
|
||||
test_pytorch_cuda && log "PyTorch CUDA compatibility confirmed" || log "WARNING: PyTorch CUDA compatibility issues detected"
|
||||
}
|
||||
|
||||
# Function to determine optimal Sage Attention strategy for mixed GPUs
|
||||
determine_sage_strategy() {
|
||||
local strategy=""
|
||||
if [ "$DETECTED_RTX20" = "true" ]; then
|
||||
if [ "$DETECTED_RTX30" = "true" ] || [ "$DETECTED_RTX40" = "true" ] || [ "$DETECTED_RTX50" = "true" ]; then
|
||||
strategy="mixed_with_rtx20"
|
||||
log "Mixed GPU setup detected with RTX 20 series - using compatibility mode"
|
||||
else
|
||||
strategy="rtx20_only"
|
||||
log "RTX 20 series only detected"
|
||||
fi
|
||||
elif [ "$DETECTED_RTX50" = "true" ]; then
|
||||
strategy="rtx50_capable"
|
||||
log "RTX 50 series detected - using latest optimizations"
|
||||
elif [ "$DETECTED_RTX40" = "true" ] || [ "$DETECTED_RTX30" = "true" ]; then
|
||||
strategy="rtx30_40_optimized"
|
||||
log "RTX 30/40 series detected - using standard optimizations"
|
||||
else
|
||||
strategy="fallback"
|
||||
log "Unknown or unsupported GPU configuration - using fallback"
|
||||
fi
|
||||
export SAGE_STRATEGY=$strategy
|
||||
local s=""
|
||||
if [ "${DET_TURING:-false}" = "true" ]; then
|
||||
if [ "${DET_AMP80:-false}" = "true" ] || [ "${DET_AMP86:-false}" = "true" ] || [ "${DET_AMP87:-false}" = "true" ] || [ "${DET_ADA:-false}" = "true" ] || [ "${DET_HOPPER:-false}" = "true" ] || [ "${DET_BW12:-false}" = "true" ] || [ "${DET_BW10:-false}" = "true" ]; then
|
||||
s="mixed_with_turing"; log "Mixed rig including Turing - using compatibility mode"
|
||||
else s="turing_only"; log "Turing-only rig detected"; fi
|
||||
elif [ "${DET_BW12:-false}" = "true" ] || [ "${DET_BW10:-false}" = "true" ]; then s="blackwell_capable"; log "Blackwell detected - using latest optimizations"
|
||||
elif [ "${DET_HOPPER:-false}" = "true" ]; then s="hopper_capable"; log "Hopper detected - using modern optimizations"
|
||||
elif [ "${DET_ADA:-false}" = "true" ] || [ "${DET_AMP86:-false}" = "true" ] || [ "${DET_AMP87:-false}" = "true" ] || [ "${DET_AMP80:-false}" = "true" ]; then
|
||||
s="ampere_ada_optimized"; log "Ampere/Ada detected - using standard optimizations"
|
||||
else s="fallback"; log "Unknown configuration - using fallback"; fi
|
||||
export SAGE_STRATEGY=$s
|
||||
}
|
||||
|
||||
# Function to install appropriate Triton version based on strategy
|
||||
install_triton_version() {
|
||||
case "$SAGE_STRATEGY" in
|
||||
"mixed_with_rtx20"|"rtx20_only")
|
||||
log "Installing Triton 3.2.0 for broader compatibility on Turing-era GPUs"
|
||||
python -m pip install --user --force-reinstall "triton==3.2.0" || {
|
||||
log "WARNING: Failed to pin Triton 3.2.0, trying latest"
|
||||
python -m pip install --user --force-reinstall triton || true
|
||||
}
|
||||
"mixed_with_turing"|"turing_only")
|
||||
log "Installing Triton 3.2.0 for Turing compatibility"
|
||||
python -m pip install --user --force-reinstall "triton==3.2.0" || python -m pip install --user --force-reinstall triton || true
|
||||
;;
|
||||
"rtx50_capable")
|
||||
log "Installing latest Triton for Blackwell/RTX 50"
|
||||
python -m pip install --user --force-reinstall triton || \
|
||||
python -m pip install --user --force-reinstall --pre triton || {
|
||||
log "WARNING: Latest Triton install failed, falling back to >=3.2.0"
|
||||
python -m pip install --user --force-reinstall "triton>=3.2.0" || true
|
||||
}
|
||||
"blackwell_capable"|"hopper_capable")
|
||||
log "Installing latest Triton for Hopper/Blackwell"
|
||||
python -m pip install --user --force-reinstall triton || python -m pip install --user --force-reinstall --pre triton || python -m pip install --user --force-reinstall "triton>=3.2.0" || true
|
||||
;;
|
||||
*)
|
||||
log "Installing latest stable Triton"
|
||||
python -m pip install --user --force-reinstall triton || {
|
||||
log "WARNING: Triton installation failed, continuing without"
|
||||
return 1
|
||||
}
|
||||
python -m pip install --user --force-reinstall triton || { log "WARNING: Triton installation failed"; return 1; }
|
||||
;;
|
||||
esac
|
||||
}
|
||||
|
||||
# Function to build Sage Attention with architecture-specific optimizations
|
||||
build_sage_attention_mixed() {
|
||||
log "Building Sage Attention for current GPU environment..."
|
||||
mkdir -p "$SAGE_ATTENTION_DIR"
|
||||
cd "$SAGE_ATTENTION_DIR"
|
||||
log "Building Sage Attention..."
|
||||
mkdir -p "$SAGE_ATTENTION_DIR"; cd "$SAGE_ATTENTION_DIR"
|
||||
|
||||
# Compute capability mapping for TORCH_CUDA_ARCH_LIST:
|
||||
# Turing = 7.5, Ampere = 8.6, Ada = 8.9, Blackwell (RTX 50) = 10.0
|
||||
# See NVIDIA Blackwell guide (sm_100/compute_100 ~ 10.0) and PyTorch arch list semantics. [doc refs in text]
|
||||
local cuda_arch_list=""
|
||||
[ "$DETECTED_RTX20" = "true" ] && cuda_arch_list="${cuda_arch_list}7.5;"
|
||||
[ "$DETECTED_RTX30" = "true" ] && cuda_arch_list="${cuda_arch_list}8.6;"
|
||||
[ "$DETECTED_RTX40" = "true" ] && cuda_arch_list="${cuda_arch_list}8.9;"
|
||||
[ "$DETECTED_RTX50" = "true" ] && cuda_arch_list="${cuda_arch_list}10.0;"
|
||||
cuda_arch_list=${cuda_arch_list%;}
|
||||
|
||||
export TORCH_CUDA_ARCH_LIST="$cuda_arch_list"
|
||||
local arch_list="${SAGE_ARCH_LIST_OVERRIDE:-$(compute_arch_list_from_torch)}"
|
||||
if [ -z "$arch_list" ]; then
|
||||
local tmp=""
|
||||
[ "${DET_TURING:-false}" = "true" ] && tmp="${tmp}7.5;"
|
||||
[ "${DET_AMP80:-false}" = "true" ] && tmp="${tmp}8.0;"
|
||||
[ "${DET_AMP86:-false}" = "true" ] && tmp="${tmp}8.6;"
|
||||
[ "${DET_AMP87:-false}" = "true" ] && tmp="${tmp}8.7;"
|
||||
[ "${DET_ADA:-false}" = "true" ] && tmp="${tmp}8.9;"
|
||||
[ "${DET_HOPPER:-false}" = "true" ] && tmp="${tmp}9.0;"
|
||||
[ "${DET_BW10:-false}" = "true" ] && tmp="${tmp}10.0;"
|
||||
[ "${DET_BW12:-false}" = "true" ] && tmp="${tmp}12.0;"
|
||||
arch_list="${tmp%;}"
|
||||
fi
|
||||
export TORCH_CUDA_ARCH_LIST="$arch_list"
|
||||
log "Set TORCH_CUDA_ARCH_LIST=$TORCH_CUDA_ARCH_LIST"
|
||||
|
||||
case "$SAGE_STRATEGY" in
|
||||
"mixed_with_rtx20"|"rtx20_only")
|
||||
log "Cloning SageAttention v1.0 for RTX 20 series compatibility"
|
||||
if [ -d "SageAttention/.git" ]; then
|
||||
cd SageAttention
|
||||
git fetch --depth 1 origin || return 1
|
||||
git checkout v1.0 2>/dev/null || git checkout -b v1.0 origin/v1.0 || return 1
|
||||
git reset --hard origin/v1.0 || return 1
|
||||
else
|
||||
rm -rf SageAttention
|
||||
git clone --depth 1 https://github.com/thu-ml/SageAttention.git -b v1.0 || return 1
|
||||
cd SageAttention
|
||||
fi
|
||||
"mixed_with_turing"|"turing_only")
|
||||
log "Cloning SageAttention v1.0 for Turing"
|
||||
if [ -d "SageAttention/.git" ]; then cd SageAttention; git fetch --depth 1 origin || return 1; git checkout v1.0 2>/dev/null || git checkout -b v1.0 origin/v1.0 || return 1; git reset --hard origin/v1.0 || return 1
|
||||
else rm -rf SageAttention; git clone --depth 1 https://github.com/thu-ml/SageAttention.git -b v1.0 || return 1; cd SageAttention; fi
|
||||
;;
|
||||
*)
|
||||
log "Cloning latest SageAttention for modern GPUs"
|
||||
if [ -d "SageAttention/.git" ]; then
|
||||
cd SageAttention
|
||||
git fetch --depth 1 origin || return 1
|
||||
git reset --hard origin/main || return 1
|
||||
else
|
||||
rm -rf SageAttention
|
||||
git clone --depth 1 https://github.com/thu-ml/SageAttention.git || return 1
|
||||
cd SageAttention
|
||||
fi
|
||||
log "Cloning latest SageAttention"
|
||||
if [ -d "SageAttention/.git" ]; then cd SageAttention; git fetch --depth 1 origin || return 1; git reset --hard origin/main || return 1
|
||||
else rm -rf SageAttention; git clone --depth 1 https://github.com/thu-ml/SageAttention.git || return 1; cd SageAttention; fi
|
||||
;;
|
||||
esac
|
||||
|
||||
log "Building SageAttention (no-build-isolation) ..."
|
||||
if MAX_JOBS=$(nproc) python -m pip install --user --no-build-isolation .; then
|
||||
echo "$SAGE_STRATEGY" > "$SAGE_ATTENTION_BUILT_FLAG"
|
||||
log "SageAttention built successfully for strategy: $SAGE_STRATEGY"
|
||||
cd "$BASE_DIR"
|
||||
return 0
|
||||
[ "${SAGE_VERBOSE_BUILD:-0}" = "1" ] && export TORCH_CPP_BUILD_VERBOSE=1
|
||||
local jobs; jobs="$(decide_build_jobs)"
|
||||
log "Using MAX_JOBS=${jobs} for SageAttention build"
|
||||
|
||||
# Set MAX_JOBS only for this build call to avoid leaking globally
|
||||
if MAX_JOBS="${jobs}" python -m pip install --user --no-build-isolation .; then
|
||||
echo "$SAGE_STRATEGY|$TORCH_CUDA_ARCH_LIST" > "$SAGE_ATTENTION_BUILT_FLAG"
|
||||
log "SageAttention built successfully"
|
||||
cd "$BASE_DIR"; return 0
|
||||
else
|
||||
log "ERROR: SageAttention build failed"
|
||||
cd "$BASE_DIR"
|
||||
return 1
|
||||
cd "$BASE_DIR"; return 1
|
||||
fi
|
||||
}
|
||||
|
||||
# Function to check if current build matches detected GPUs
|
||||
needs_rebuild() {
|
||||
if [ ! -f "$SAGE_ATTENTION_BUILT_FLAG" ]; then
|
||||
return 0
|
||||
fi
|
||||
local built_strategy
|
||||
built_strategy=$(cat "$SAGE_ATTENTION_BUILT_FLAG" 2>/dev/null || echo "unknown")
|
||||
if [ "$built_strategy" != "$SAGE_STRATEGY" ]; then
|
||||
log "GPU configuration changed (was: $built_strategy, now: $SAGE_STRATEGY) - rebuild needed"
|
||||
return 0
|
||||
fi
|
||||
if [ ! -f "$SAGE_ATTENTION_BUILT_FLAG" ]; then return 0; fi
|
||||
local x; x=$(cat "$SAGE_ATTENTION_BUILT_FLAG" 2>/dev/null || echo "")
|
||||
local prev_strategy="${x%%|*}"; local prev_arch="${x#*|}"
|
||||
if [ "$prev_strategy" != "$SAGE_STRATEGY" ] || [ "$prev_arch" != "$TORCH_CUDA_ARCH_LIST" ]; then return 0; fi
|
||||
return 1
|
||||
}
|
||||
|
||||
# Function to check if SageAttention is working
|
||||
test_sage_attention() {
|
||||
python -c "
|
||||
import sys
|
||||
try:
|
||||
import sageattention
|
||||
print('[TEST] SageAttention import: SUCCESS')
|
||||
try:
|
||||
v = getattr(sageattention, '__version__', None)
|
||||
if v: print(f'[TEST] Version: {v}')
|
||||
except:
|
||||
pass
|
||||
sys.exit(0)
|
||||
import sageattention; print('[TEST] SageAttention import: SUCCESS')
|
||||
v=getattr(sageattention,'__version__',None)
|
||||
if v: print(f'[TEST] Version: {v}'); sys.exit(0)
|
||||
except ImportError as e:
|
||||
print(f'[TEST] SageAttention import: FAILED - {e}')
|
||||
sys.exit(1)
|
||||
print(f'[TEST] SageAttention import: FAILED - {e}'); sys.exit(1)
|
||||
except Exception as e:
|
||||
print(f'[TEST] SageAttention test: ERROR - {e}')
|
||||
sys.exit(1)
|
||||
print(f'[TEST] SageAttention test: ERROR - {e}'); sys.exit(1)
|
||||
" 2>/dev/null
|
||||
}
|
||||
|
||||
# Main GPU detection and SageAttention setup
|
||||
setup_sage_attention() {
|
||||
# Export build-visible status flags
|
||||
export SAGE_ATTENTION_BUILT=0
|
||||
export SAGE_ATTENTION_AVAILABLE=0
|
||||
|
||||
if ! detect_gpu_generations; then
|
||||
log "No GPUs detected, skipping SageAttention setup"
|
||||
return 0
|
||||
fi
|
||||
|
||||
export SAGE_ATTENTION_BUILT=0 SAGE_ATTENTION_AVAILABLE=0
|
||||
if ! detect_gpu_generations; then log "No GPUs detected, skipping SageAttention setup"; return 0; fi
|
||||
determine_sage_strategy
|
||||
|
||||
# Resolve arch list early
|
||||
export TORCH_CUDA_ARCH_LIST="${SAGE_ARCH_LIST_OVERRIDE:-$(compute_arch_list_from_torch)}"
|
||||
if [ -z "$TORCH_CUDA_ARCH_LIST" ]; then
|
||||
local tmp=""
|
||||
[ "${DET_TURING:-false}" = "true" ] && tmp="${tmp}7.5;"
|
||||
[ "${DET_AMP80:-false}" = "true" ] && tmp="${tmp}8.0;"
|
||||
[ "${DET_AMP86:-false}" = "true" ] && tmp="${tmp}8.6;"
|
||||
[ "${DET_AMP87:-false}" = "true" ] && tmp="${tmp}8.7;"
|
||||
[ "${DET_ADA:-false}" = "true" ] && tmp="${tmp}8.9;"
|
||||
[ "${DET_HOPPER:-false}" = "true" ] && tmp="${tmp}9.0;"
|
||||
[ "${DET_BW10:-false}" = "true" ] && tmp="${tmp}10.0;"
|
||||
[ "${DET_BW12:-false}" = "true" ] && tmp="${tmp}12.0;"
|
||||
export TORCH_CUDA_ARCH_LIST="${tmp%;}"
|
||||
fi
|
||||
log "Resolved TORCH_CUDA_ARCH_LIST=$TORCH_CUDA_ARCH_LIST"
|
||||
|
||||
if needs_rebuild || ! test_sage_attention; then
|
||||
log "Building SageAttention..."
|
||||
if install_triton_version && build_sage_attention_mixed && test_sage_attention; then
|
||||
export SAGE_ATTENTION_BUILT=1
|
||||
export SAGE_ATTENTION_AVAILABLE=1
|
||||
export SAGE_ATTENTION_BUILT=1 SAGE_ATTENTION_AVAILABLE=1
|
||||
log "SageAttention is built; set FORCE_SAGE_ATTENTION=1 to enable it at startup"
|
||||
else
|
||||
export SAGE_ATTENTION_BUILT=0
|
||||
export SAGE_ATTENTION_AVAILABLE=0
|
||||
export SAGE_ATTENTION_BUILT=0 SAGE_ATTENTION_AVAILABLE=0
|
||||
log "WARNING: SageAttention is not available after build attempt"
|
||||
fi
|
||||
else
|
||||
export SAGE_ATTENTION_BUILT=1
|
||||
export SAGE_ATTENTION_AVAILABLE=1
|
||||
log "SageAttention already built and importable for current GPU configuration"
|
||||
export SAGE_ATTENTION_BUILT=1 SAGE_ATTENTION_AVAILABLE=1
|
||||
log "SageAttention already built and importable"
|
||||
fi
|
||||
}
|
||||
|
||||
# If running as root, handle permissions and user mapping
|
||||
# --- root to runtime user ---
|
||||
if [ "$(id -u)" = "0" ]; then
|
||||
if [ ! -f "$PERMISSIONS_SET_FLAG" ]; then
|
||||
log "Setting up user permissions..."
|
||||
if getent group "${PGID}" >/dev/null; then
|
||||
EXISTING_GRP="$(getent group "${PGID}" | cut -d: -f1)"
|
||||
usermod -g "${EXISTING_GRP}" "${APP_USER}" || true
|
||||
APP_GROUP="${EXISTING_GRP}"
|
||||
else
|
||||
groupmod -o -g "${PGID}" "${APP_GROUP}" || true
|
||||
fi
|
||||
EXISTING_GRP="$(getent group "${PGID}" | cut -d: -f1)"; usermod -g "${EXISTING_GRP}" "${APP_USER}" || true; APP_GROUP="${EXISTING_GRP}"
|
||||
else groupmod -o -g "${PGID}" "${APP_GROUP}" || true; fi
|
||||
usermod -o -u "${PUID}" "${APP_USER}" || true
|
||||
mkdir -p "/home/${APP_USER}"
|
||||
for d in "$BASE_DIR" "/home/$APP_USER"; do
|
||||
[ -e "$d" ] && chown -R "${APP_USER}:${APP_GROUP}" "$d" || true
|
||||
done
|
||||
for d in "$BASE_DIR" "/home/$APP_USER"; do [ -e "$d" ] && chown -R "${APP_USER}:${APP_GROUP}" "$d" || true; done
|
||||
|
||||
readarray -t PY_PATHS < <(python - <<'PY'
|
||||
import sys, sysconfig, os, datetime
|
||||
def log(msg):
|
||||
import datetime
|
||||
ts = datetime.datetime.now().strftime("%H:%M:%S")
|
||||
print(f"[bootstrap:python {ts}] {msg}", file=sys.stderr, flush=True)
|
||||
def log(m): print(f"[bootstrap:python {datetime.datetime.now().strftime('%H:%M:%S')}] {m}", file=sys.stderr, flush=True)
|
||||
log("Determining writable Python install targets via sysconfig.get_paths()")
|
||||
keys = ("purelib","platlib","scripts","include","platinclude","data")
|
||||
paths = sysconfig.get_paths()
|
||||
for k in keys:
|
||||
v = paths.get(k)
|
||||
if v:
|
||||
print(v)
|
||||
log(f"emit {k} -> {v}")
|
||||
d = paths.get("data")
|
||||
for k in ("purelib","platlib","scripts","include","platinclude","data"):
|
||||
v = sysconfig.get_paths().get(k)
|
||||
if v: print(v); log(f"emit {k} -> {v}")
|
||||
d = sysconfig.get_paths().get("data")
|
||||
if d:
|
||||
share = os.path.join(d, "share")
|
||||
man1 = os.path.join(share, "man", "man1")
|
||||
print(share)
|
||||
print(man1)
|
||||
log(f"emit wheel data dirs -> {share}, {man1}")
|
||||
log("Finished emitting target directories")
|
||||
share=os.path.join(d,"share"); man1=os.path.join(share,"man","man1")
|
||||
print(share); print(man1); log(f"emit wheel data dirs -> {share}, {man1}")
|
||||
PY
|
||||
)
|
||||
for d in "${PY_PATHS[@]}"; do
|
||||
case "$d" in
|
||||
/usr/local|/usr/local/*)
|
||||
mkdir -p "$d" || true
|
||||
chown -R "${APP_USER}:${APP_GROUP}" "$d" || true
|
||||
chmod -R u+rwX,g+rwX "$d" || true
|
||||
;;
|
||||
/usr/local|/usr/local/*) mkdir -p "$d" || true; chown -R "${APP_USER}:${APP_GROUP}" "$d" || true; chmod -R u+rwX,g+rwX "$d" || true ;;
|
||||
*) : ;;
|
||||
esac
|
||||
done
|
||||
touch "$PERMISSIONS_SET_FLAG"
|
||||
chown "${APP_USER}:${APP_GROUP}" "$PERMISSIONS_SET_FLAG"
|
||||
touch "$PERMISSIONS_SET_FLAG"; chown "${APP_USER}:${APP_GROUP}" "$PERMISSIONS_SET_FLAG"
|
||||
log "User permissions configured"
|
||||
else
|
||||
log "User permissions already configured, skipping..."
|
||||
fi
|
||||
else log "User permissions already configured, skipping..."; fi
|
||||
exec runuser -u "${APP_USER}" -- "$0" "$@"
|
||||
fi
|
||||
|
||||
# Setup SageAttention for detected GPU configuration
|
||||
# --- SageAttention setup ---
|
||||
setup_sage_attention
|
||||
|
||||
# Ensure ComfyUI-Manager exists or update it (shallow)
|
||||
# --- ComfyUI-Manager sync ---
|
||||
if [ -d "$CUSTOM_NODES_DIR/ComfyUI-Manager/.git" ]; then
|
||||
log "Updating ComfyUI-Manager in $CUSTOM_NODES_DIR/ComfyUI-Manager"
|
||||
git -C "$CUSTOM_NODES_DIR/ComfyUI-Manager" fetch --depth 1 origin || true
|
||||
git -C "$CUSTOM_NODES_DIR/ComfyUI-Manager" reset --hard origin/HEAD || true
|
||||
git -C "$CUSTOM_NODES_DIR/ComfyUI-Manager" clean -fdx || true
|
||||
log "Updating ComfyUI-Manager"; git -C "$CUSTOM_NODES_DIR/ComfyUI-Manager" fetch --depth 1 origin || true
|
||||
git -C "$CUSTOM_NODES_DIR/ComfyUI-Manager" reset --hard origin/HEAD || true; git -C "$CUSTOM_NODES_DIR/ComfyUI-Manager" clean -fdx || true
|
||||
elif [ ! -d "$CUSTOM_NODES_DIR/ComfyUI-Manager" ]; then
|
||||
log "Installing ComfyUI-Manager into $CUSTOM_NODES_DIR/ComfyUI-Manager"
|
||||
git clone --depth 1 https://github.com/ltdrdata/ComfyUI-Manager.git "$CUSTOM_NODES_DIR/ComfyUI-Manager" || true
|
||||
log "Installing ComfyUI-Manager"; git clone --depth 1 https://github.com/ltdrdata/ComfyUI-Manager.git "$CUSTOM_NODES_DIR/ComfyUI-Manager" || true
|
||||
fi
|
||||
|
||||
# User-site PATHs for --user installs (custom nodes)
|
||||
# --- PATH/PYTHONPATH ---
|
||||
export PATH="$HOME/.local/bin:$PATH"
|
||||
pyver="$(python -c 'import sys; print(f"{sys.version_info.major}.{sys.version_info.minor}")')"
|
||||
export PYTHONPATH="$HOME/.local/lib/python${pyver}/site-packages:${PYTHONPATH:-}"
|
||||
|
||||
# First-run driven auto-install of custom node deps
|
||||
# --- first-run install of custom_nodes ---
|
||||
if [ ! -f "$FIRST_RUN_FLAG" ] || [ "${COMFY_FORCE_INSTALL:-0}" = "1" ]; then
|
||||
if [ "${COMFY_AUTO_INSTALL:-1}" = "1" ]; then
|
||||
log "First run detected or forced; scanning custom nodes for requirements..."
|
||||
# requirements*.txt
|
||||
log "First run or forced; installing custom node dependencies..."
|
||||
while IFS= read -r -d '' req; do
|
||||
log "python -m pip install --user --upgrade -r $req"
|
||||
python -m pip install --no-cache-dir --user --upgrade --upgrade-strategy only-if-needed -r "$req" || true
|
||||
done < <(find "$CUSTOM_NODES_DIR" -maxdepth 3 -type f \( -iname 'requirements.txt' -o -iname 'requirements-*.txt' -o -path '*/requirements/*.txt' \) -print0)
|
||||
|
||||
# pyproject.toml (exclude ComfyUI-Manager)
|
||||
while IFS= read -r -d '' pjt; do
|
||||
d="$(dirname "$pjt")"
|
||||
log "python -m pip install --user . in $d"
|
||||
d="$(dirname "$pjt")"; log "python -m pip install --user . in $d"
|
||||
(cd "$d" && python -m pip install --no-cache-dir --user .) || true
|
||||
done < <(find "$CUSTOM_NODES_DIR" -maxdepth 2 -type f -iname 'pyproject.toml' -not -path '*/ComfyUI-Manager/*' -print0)
|
||||
|
||||
python -m pip check || true
|
||||
else
|
||||
log "COMFY_AUTO_INSTALL=0; skipping dependency install on first run"
|
||||
fi
|
||||
else log "COMFY_AUTO_INSTALL=0; skipping dependency install"; fi
|
||||
touch "$FIRST_RUN_FLAG"
|
||||
else
|
||||
log "Not first run; skipping custom_nodes dependency install"
|
||||
fi
|
||||
|
||||
# Build ComfyUI command with SageAttention usage controlled only by FORCE_SAGE_ATTENTION
|
||||
# --- launch ComfyUI ---
|
||||
COMFYUI_ARGS=""
|
||||
if [ "${FORCE_SAGE_ATTENTION:-0}" = "1" ]; then
|
||||
if test_sage_attention; then
|
||||
COMFYUI_ARGS="--use-sage-attention"
|
||||
log "Starting ComfyUI with SageAttention enabled by environment (FORCE_SAGE_ATTENTION=1)"
|
||||
else
|
||||
log "WARNING: FORCE_SAGE_ATTENTION=1 but SageAttention import failed; starting without"
|
||||
fi
|
||||
if test_sage_attention; then COMFYUI_ARGS="--use-sage-attention"; log "Starting ComfyUI with SageAttention (FORCE_SAGE_ATTENTION=1)"
|
||||
else log "WARNING: FORCE_SAGE_ATTENTION=1 but import failed; starting without"; fi
|
||||
else
|
||||
if [ "${SAGE_ATTENTION_AVAILABLE:-0}" = "1" ]; then
|
||||
log "SageAttention is built; set FORCE_SAGE_ATTENTION=1 to enable it at startup"
|
||||
else
|
||||
log "SageAttention not available; starting without it"
|
||||
fi
|
||||
if [ "${SAGE_ATTENTION_AVAILABLE:-0}" = "1" ]; then log "SageAttention is built; set FORCE_SAGE_ATTENTION=1 to enable"
|
||||
else log "SageAttention not available; starting without it"; fi
|
||||
fi
|
||||
|
||||
cd "$BASE_DIR"
|
||||
|
||||
# Handle both direct execution and passed arguments
|
||||
if [ $# -eq 0 ]; then
|
||||
exec python main.py --listen 0.0.0.0 $COMFYUI_ARGS
|
||||
if [ $# -eq 0 ]; then exec python main.py --listen 0.0.0.0 $COMFYUI_ARGS
|
||||
else
|
||||
if [ "$1" = "python" ] && [ "${2:-}" = "main.py" ]; then
|
||||
shift 2
|
||||
exec python main.py $COMFYUI_ARGS "$@"
|
||||
else
|
||||
exec "$@"
|
||||
fi
|
||||
if [ "$1" = "python" ] && [ "${2:-}" = "main.py" ]; then shift 2; exec python main.py $COMFYUI_ARGS "$@"
|
||||
else exec "$@"; fi
|
||||
fi
|
||||
|
||||
Loading…
Reference in New Issue
Block a user