b300/fix2.sh

#!/usr/bin/env bash
# Second repair. --no-deps left torch 2.12.0 without a matching NCCL
# (undefined symbol: ncclCommResume). Install torch WITH its dependency set from
# the cu130 index so the whole nvidia stack is consistent, then re-verify.
set -uo pipefail
RT=/KRAFTON/WORKSPACE/wbl-workspace/posttraining-2606/.b300-runtime
PY=$RT/bin/python
LOG=$RT/restore.log
log(){ echo "[$(date '+%F %H:%M:%S')] FIX2 $*" | tee -a "$LOG"; }
log "reinstalling torch 2.12.0+cu130 WITH deps"
$PY -m pip install --no-cache-dir --force-reinstall \
  torch==2.12.0 --index-url https://download.pytorch.org/whl/cu130 >>"$LOG" 2>&1
log "torch rc=$?"
log "=== verification ==="
$PY - <<'PY' 2>&1 | tee -a "$LOG"
import importlib
for m in ["torch","transformer_engine","megatron.core","megatron.bridge","transformers","mamba_ssm","causal_conv1d"]:
    try:
        mod=importlib.import_module(m)
        print(f"  OK   {m:20s} {getattr(mod,'__version__','')}")
    except Exception as e:
        print(f"  MISS {m:20s} {type(e).__name__}: {str(e)[:70]}")
PY
log "FIX2 DONE"