fridge_m setup: chunked_runner + run_gen_chunk + check_mismatch + push_watchdog

fridge_m/_setup/scripts/extend_chunks_89.sh

@@ -0,0 +1,40 @@
+#!/bin/bash
+# After chunk 7 + its final flush finish, run chunks 8 and 9 (sim+gen each),
+# then do another final flush to push the additional ~100 trajs.
+#
+# Sequencing rationale: chunk 7 gen is using all 4 GPUs; co-launching chunks 8
+# sim now would force VRAM sharing (gen 21.2 GB + sim 1.5 GB per GPU = 22.7 GB
+# on a 24 GB 3090, too close to OOM). Wait for chunk 7 to fully complete, then
+# sequential chunks 8 and 9 the same way chunks 1-6 ran.
+set -uo pipefail
+
+CHUNK7_RUNNER_PID=${CHUNK7_RUNNER_PID:-1546061}
+POST7_PID=${POST7_PID:-1546543}
+
+echo "[ext] waiting for chunk-7 chunked_runner pid=$CHUNK7_RUNNER_PID ..."
+while kill -0 "$CHUNK7_RUNNER_PID" 2>/dev/null; do sleep 30; done
+echo "[ext] chunked_runner (chunk 7) exited at $(date -Iseconds)"
+
+echo "[ext] waiting for post_chunk7_flush pid=$POST7_PID (final HF push) ..."
+while kill -0 "$POST7_PID" 2>/dev/null; do sleep 15; done
+echo "[ext] post_chunk7_flush exited at $(date -Iseconds)"
+
+# Sanity: confirm no GPU users remain
+sleep 5
+echo "[ext] residual GPU users (should be empty):"
+pgrep -af '[r]un_pipeline_no_clip\.py|[i]nference_long_video\.py|[d]epth_methods\.py' 2>/dev/null || true
+
+# Archive the chunk-7 runner log
+mv /scr/ravenh/fridge_m/logs/chunked_runner.log /scr/ravenh/fridge_m/logs/chunked_runner.log.chunk7done 2>/dev/null || true
+
+echo "[ext] launching chunked_runner for chunks 8..9 at $(date -Iseconds)"
+START_CHUNK=8 NUM_CHUNKS=9 SAMPLES_PER_HOUSE=3 SEED_BASE=30000 GPUS="0,1,2,3" STOP_AT_TRAJS=100000 \
+  bash /scr/ravenh/fridge_m/scripts/chunked_runner.sh
+echo "[ext] chunked_runner (chunks 8..9) exited at $(date -Iseconds)"
+
+# Re-run the final flush to push chunks 7+8+9 (chunk 7 was already pushed in post7).
+echo "[ext] final HF flush at $(date -Iseconds)"
+source /svl/u/ravenh/miniconda3/etc/profile.d/conda.sh
+conda activate mlspaces
+python -u /scr/ravenh/fridge_m/scripts/final_flush.py 2>&1 | tee -a /scr/ravenh/fridge_m/logs/final_flush.log
+echo "[ext] DONE at $(date -Iseconds)"

fridge_m/_setup/scripts/extend_chunks_89_simfirst.sh

@@ -0,0 +1,111 @@
+#!/bin/bash
+# After chunk 7 + its final flush finish, do:
+#   1. chunk 8 sim
+#   2. chunk 9 sim    (only after chunk 8 sim done; both use 4 GPUs)
+#   3. chunk 8 gen
+#   4. chunk 9 gen
+#   5. final HF flush
+#
+# I.e. all sim work first, then all gen work. Different from the chunked_runner
+# default of sim+gen interleaved per chunk, per user request.
+set -uo pipefail
+
+CHUNK7_RUNNER_PID=${CHUNK7_RUNNER_PID:-1546061}
+POST7_PID=${POST7_PID:-1546543}
+REPO_MOLMO=/svl/u/ravenh/renderscale/molmospaces
+ROOT=/scr/ravenh/fridge_m
+SCRIPTS=$ROOT/scripts
+SIM_CHUNKS=$ROOT/sim_chunks
+GEN_CHUNKS=$ROOT/gen_chunks
+SCRATCH_GEN=$ROOT/_workdir_gen
+LOGS=$ROOT/logs
+GPUS=${GPUS:-0,1,2,3}
+SAMPLES_PER_HOUSE=${SAMPLES_PER_HOUSE:-3}
+SEED_BASE=${SEED_BASE:-30000}
+
+source /svl/u/ravenh/miniconda3/etc/profile.d/conda.sh
+conda activate mlspaces
+
+# data_gen_fridge_m.sh expects cwd=$REPO_MOLMO (relative camera_extrinsics path).
+cd "$REPO_MOLMO"
+
+# ---- Phase 1: wait for chunk 7 + post7-flush to fully exit ----
+echo "[ext-sf] waiting for chunk-7 chunked_runner pid=$CHUNK7_RUNNER_PID ..."
+while kill -0 "$CHUNK7_RUNNER_PID" 2>/dev/null; do sleep 30; done
+echo "[ext-sf] chunked_runner (chunk 7) exited at $(date -Iseconds)"
+
+echo "[ext-sf] waiting for post_chunk7_flush pid=$POST7_PID (final HF push) ..."
+while kill -0 "$POST7_PID" 2>/dev/null; do sleep 15; done
+echo "[ext-sf] post_chunk7_flush exited at $(date -Iseconds)"
+
+sleep 5
+echo "[ext-sf] residual GPU users (should be empty):"
+pgrep -af '[r]un_pipeline_no_clip\.py|[i]nference_long_video\.py|[d]epth_methods\.py' 2>/dev/null || true
+
+# ---- Phase 2: chunk 8 sim ----
+launch_sim_chunk() {
+  local n=$1
+  local chunk_dir="$SIM_CHUNKS/chunk_$(printf '%03d' $n)"
+  mkdir -p "$chunk_dir"
+  echo "[ext-sf] phase: chunk $n sim -> $chunk_dir   $(date -Iseconds)"
+  declare -a PIDS=()
+  for s in 1 2 3 4; do
+    GPU=$((s - 1))
+    SEED=$((SEED_BASE + n*100 + s))
+    LOG="$LOGS/sim_chunk_$(printf '%03d' $n)_shard_${s}.log"
+    CUDA_VISIBLE_DEVICES=$GPU OUTPUT_DIR="$chunk_dir" SAMPLES_PER_HOUSE="$SAMPLES_PER_HOUSE" \
+      nohup bash "$REPO_MOLMO/scripts/datagen/data_gen_fridge_m.sh" "$s" 4 "$SEED" \
+      > "$LOG" 2>&1 &
+    PIDS+=($!)
+    echo "  shard $s/4 GPU=$GPU SEED=$SEED PID=$! log=$LOG"
+  done
+  echo "[ext-sf] waiting for chunk $n sim shards to exit..."
+  for pid in "${PIDS[@]}"; do
+    if ! wait "$pid"; then
+      echo "[ext-sf] WARN: sim shard pid=$pid exited non-zero"
+    fi
+  done
+  echo "[ext-sf] chunk $n sim done at $(date -Iseconds)"
+  echo "[ext-sf] check_mismatch on chunk $n ..."
+  if ! python "$SCRIPTS/check_mismatch.py" --sim_dir "$chunk_dir"; then
+    echo "[ext-sf] FATAL: mismatch in chunk $n; aborting before launching any more work"
+    exit 2
+  fi
+}
+
+run_gen_chunk_oneshot() {
+  local n=$1
+  local chunk_dir="$SIM_CHUNKS/chunk_$(printf '%03d' $n)"
+  local gen_dir="$GEN_CHUNKS/chunk_$(printf '%03d' $n)"
+  local scratch="$SCRATCH_GEN/chunk_$(printf '%03d' $n)"
+  mkdir -p "$gen_dir" "$scratch"
+  local count=$(ls "$chunk_dir"/episode_*_exo_camera_1_robot_rgb*.mp4 2>/dev/null | wc -l)
+  echo "[ext-sf] phase: chunk $n gen ($count trajs) using GPUs=$GPUS   $(date -Iseconds)"
+  SEQ_OFFSET=$((n*100))
+  python "$SCRIPTS/run_gen_chunk.py" \
+    --sim_src "$chunk_dir" \
+    --out_dir "$gen_dir" \
+    --scratch "$scratch" \
+    --gpus "$GPUS" \
+    --seq_offset "$SEQ_OFFSET" 2>&1 | stdbuf -oL tee "$LOGS/gen_chunk_$(printf '%03d' $n).log" || \
+    echo "[ext-sf] WARN: gen chunk $n exited non-zero (some trajs may have failed; continuing)"
+  echo "[ext-sf] chunk $n gen done at $(date -Iseconds)"
+  rm -rf "$scratch/_inf_tmp" 2>/dev/null || true
+}
+
+# Phase 2: chunk 8 sim
+launch_sim_chunk 8
+
+# Phase 3: chunk 9 sim (sequential — must wait for chunk 8 sim before launching 9 sim, so the 4 GPUs free up)
+launch_sim_chunk 9
+
+# Phase 4: chunk 8 gen
+run_gen_chunk_oneshot 8
+
+# Phase 5: chunk 9 gen
+run_gen_chunk_oneshot 9
+
+# Phase 6: final HF flush (pushes chunks 1..9 sim+gen+h5)
+echo "[ext-sf] final HF flush at $(date -Iseconds)"
+python -u "$SCRIPTS/final_flush.py" 2>&1 | tee -a "$LOGS/final_flush.log"
+echo "[ext-sf] DONE at $(date -Iseconds)"

fridge_m/_setup/scripts/post_chunk7_flush.sh

@@ -0,0 +1,31 @@
+#!/bin/bash
+# Waits for chunked_runner to fully exit (chunk 7 is the only chunk in this
+# run, so when it exits the loop is done), then re-runs the final HF flush.
+set -uo pipefail
+
+RUNNER_PID=${RUNNER_PID:-1546061}
+
+echo "[post7] waiting for chunked_runner pid=$RUNNER_PID to exit ..."
+while kill -0 "$RUNNER_PID" 2>/dev/null; do
+  sleep 30
+done
+echo "[post7] chunked_runner exited at $(date -Iseconds); waiting 10s to settle..."
+sleep 10
+
+# Confirm chunk 7 produced trajs
+n=$(ls /scr/ravenh/fridge_m/sim_chunks/chunk_007/episode_*_exo_camera_1_robot_rgb*.mp4 2>/dev/null | wc -l)
+echo "[post7] chunk_007 finalized sim trajs: $n"
+if [ "$n" = "0" ]; then
+  echo "[post7] WARN: chunk 7 produced 0 trajs; nothing new to flush. Exiting."
+  exit 1
+fi
+
+g=$(ls /scr/ravenh/fridge_m/gen_chunks/chunk_007/generated_episode_*_exo_camera_1*.mp4 2>/dev/null | wc -l)
+echo "[post7] chunk_007 finalized gen trajs: $g"
+
+source /svl/u/ravenh/miniconda3/etc/profile.d/conda.sh
+conda activate mlspaces
+
+echo "[post7] starting final HF flush at $(date -Iseconds)"
+python -u /scr/ravenh/fridge_m/scripts/final_flush.py 2>&1 | tee -a /scr/ravenh/fridge_m/logs/final_flush.log
+echo "[post7] final flush done at $(date -Iseconds)"