Validate Puzzletron campaign configuration before launch

examples/puzzletron/configs/orchestration/execution.example.yaml

@@ -10,12 +10,24 @@ execution:
     artifact_settling_timeout_seconds: 300
     gpus_per_node: 8
   stages:
+    # CPU-only stages request no GPUs and can override the runner partition.
+    # A list lets Slurm select any eligible CPU partition.
     convert:
       strategy: single
       instances: 1
+      resource: cpu
+      partition:
+        - REPLACE_WITH_PRIMARY_CPU_SLURM_PARTITION
+        - REPLACE_WITH_ALTERNATE_CPU_SLURM_PARTITION
     tokenize_data:
       strategy: single
       instances: 1
+      resource: cpu
+      partition: REPLACE_WITH_PRIMARY_CPU_SLURM_PARTITION
+    # Final report generation is always CPU-only and accepts a partition
+    # override without resource or strategy fields.
+    final_report:
+      partition: REPLACE_WITH_PRIMARY_CPU_SLURM_PARTITION
     vllm_stats:
       strategy: sharded
       instances: 16

examples/puzzletron/configs/orchestration/qwen3p5_0p8b/runner.slurm.yaml

@@ -8,13 +8,8 @@ runner:
   slurm:
     account: REPLACE_WITH_SLURM_ACCOUNT
     partition: REPLACE_WITH_SLURM_PARTITION
-    partition_batch: REPLACE_WITH_SLURM_PARTITION
-    partition_interactive:
-    partition_cpu:
-    interactive_max_nodes: 1
     max_nodes: 1
     time_limit: "1:00:00"
-    log_dir: puzzle_runs/qwen3p5_0p8b_smoke/logs
   execution_contract:
     repository: REPLACE_WITH_WORKER_VISIBLE_MODELOPT_CHECKOUT
     venv: REPLACE_WITH_WORKER_VISIBLE_MODELOPT_VENV

examples/puzzletron/configs/orchestration/qwen_moe/execution.production.yaml

@@ -9,16 +9,16 @@ execution:
   stages:
     convert: {strategy: single, instances: 1}
     tokenize_data: {strategy: single, instances: 1}
-    # One interactive node (8 GPUs): pack 8 vLLM shards onto that single node.
-    vllm_stats: {strategy: sharded, instances: 8, partition: interactive}
-    width_importance: {strategy: single, instances: 1, partition: interactive}
+    # One node (8 GPUs): pack 8 vLLM shards onto that single node.
+    vllm_stats: {strategy: sharded, instances: 8}
+    width_importance: {strategy: single, instances: 1}
     # One gang-scheduled allocation: four 8-GPU workers, coordinator on node 0.
-    depth_importance: {strategy: persistent_pool, instances: 4, partition: batch}
-    sort: {strategy: single, instances: 1, partition: interactive}
-    sort_sanity: {strategy: single, instances: 1, partition: interactive}
-    width_sanity: {strategy: single, instances: 1, partition: interactive}
-    slicing_sanity: {strategy: single, instances: 1, partition: interactive}
-    bypass_sanity: {strategy: single, instances: 1, partition: interactive}
+    depth_importance: {strategy: persistent_pool, instances: 4}
+    sort: {strategy: single, instances: 1}
+    sort_sanity: {strategy: single, instances: 1}
+    width_sanity: {strategy: single, instances: 1}
+    slicing_sanity: {strategy: single, instances: 1}
+    bypass_sanity: {strategy: single, instances: 1}
     bypass: {strategy: single, instances: 1}
     build_library:
       strategy: single
@@ -32,7 +32,7 @@ execution:
         dp_shard: 1
         dp_replicate: 1
     # One gang-scheduled allocation: two 8-GPU workers, coordinator on node 0.
-    replacement_scoring: {strategy: persistent_pool, instances: 2, partition: interactive}
+    replacement_scoring: {strategy: persistent_pool, instances: 2}
     mip: {strategy: single, instances: 1}
     zero_shot_evaluation: {strategy: sharded, instances: 8}
     aiperf: {strategy: sharded, instances: 8}

examples/puzzletron/configs/orchestration/qwen_moe/runner.slurm.yaml

@@ -6,16 +6,10 @@ runner:
   slurm:
     # Required placeholder. Replace with the Slurm account for your site.
     account: REPLACE_WITH_SLURM_ACCOUNT
-    # Replace these generic partition names if your site uses different names.
-    partition: batch
-    partition_interactive: interactive
-    partition_batch: batch
-    # Optional. CPU/IO stages use the regular one-node partition when unset.
-    partition_cpu:
-    interactive_max_nodes: 2
+    # Replace with one partition name or a list of eligible names for this site.
+    partition: REPLACE_WITH_SLURM_PARTITION
     max_nodes: 20
     time_limit: "4:00:00"
-    log_dir: puzzle_runs/qwen-moe/logs
   execution_contract:
     # Replace with the ModelOpt checkout path visible on every worker and in the container.
     repository: REPLACE_WITH_WORKER_VISIBLE_MODELOPT_CHECKOUT

examples/puzzletron/configs/orchestration/runner.slurm.example.yaml

@@ -8,15 +8,13 @@ runner:
   slurm:
     # Required placeholder. Replace with the Slurm account for your site.
     account: REPLACE_WITH_SLURM_ACCOUNT
-    # Replace these generic partition names if your site uses different names.
-    partition: batch
-    partition_interactive: interactive
-    partition_batch: batch
-    # Optional. CPU/IO stages use the regular one-node partition when unset.
-    partition_cpu:
-    interactive_max_nodes: 2
+    # Optional default for stages without a partition override. Use one name or
+    # a list of eligible names. Slurm uses the site default when this is omitted.
+    partition:
+      - REPLACE_WITH_PRIMARY_SLURM_PARTITION
+      - REPLACE_WITH_ALTERNATE_SLURM_PARTITION
     time_limit: "4:00:00"
-    log_dir: puzzle_runs/logs
+    log_dir: logs
   execution_contract:
     # Required. Use the checkout path visible on every worker and in the container, if used.
     repository: REPLACE_WITH_WORKER_VISIBLE_MODELOPT_CHECKOUT

examples/puzzletron/configs/setup/defaults.example.yaml

@@ -19,5 +19,6 @@ infrastructure:
     slurm:
       # Add the required Slurm account for your site before using this file.
       # account: REPLACE_WITH_SLURM_ACCOUNT
-      # Optional. CPU/IO stages use the regular partition when unset.
-      partition_cpu:
+      # Optional. Use one partition name or a list of eligible names. When
+      # omitted, Slurm uses the site's default partition.
+      partition:

examples/puzzletron/docs/campaign_reports.md

@@ -1,11 +1,31 @@
 # Puzzletron Campaign Reports
 
-This page catalogs retained Puzzletron campaign reports and the status of their
-evidence. The compact [campaign report index](../reports/campaign_report_index.yaml)
+The orchestrator generates a cumulative HTML report after a campaign. Regenerate
+it without rerunning model work:
+
+```bash
+python examples/puzzletron/generate_campaign_progress_report.py \
+  --puzzle-dir /shared/puzzle_runs/my_campaign \
+  --model-name 'My model'
+```
+
+The output is
+`<puzzle-dir>/artifacts/campaign_report/campaign_report.html`. Section inputs
+and configuration fingerprints are cached under
+`<puzzle-dir>/artifacts/campaign_report/section_cache`. Use
+`--rebuild-section aiperf` to rebuild one section, or `--no-cache` to rebuild
+the whole report.
+
+This page also catalogs retained Puzzletron campaign reports and the status of
+their evidence. The compact [campaign report index](../reports/campaign_report_index.yaml)
 records each report's producer state, reproduction and support status, metadata
 origin, current-configuration relationship, and known limitations. Detailed
 run facts remain in the reports.
 
+Retained reports are self-contained HTML files and may be hundreds of MB.
+Download them and open them locally. Interpret their results together with the
+reproduction status and unresolved findings below.
+
 ## Report status
 
 | Model | Report | Producer state | Reproduction | Support | Current configuration relationship |

examples/puzzletron/docs/configuration_overrides.md

@@ -0,0 +1,20 @@
+# Experiment overrides
+
+Use command-line overrides for temporary experiment value changes. Append a
+repeatable `--override KEY=VALUE` to the orchestrator command and inspect the
+result with `--dry-run` before launch:
+
+```bash
+--override mip.runs.params-90.solver.num_solutions=4 \
+--override ++runtime_annotations.reason=capacity-check \
+--dry-run
+```
+
+Plain `KEY=VALUE` and explicit `++KEY=VALUE` both add or replace experiment
+values. The controller and GPU workers interpret these forms identically.
+Single-plus add (`+KEY=VALUE`) and delete (`~KEY`) operators are not
+supported. Put structural changes in a copied run config so they remain easy to
+review.
+
+Overrides apply only to the experiment config. Edit or copy the runner and
+execution files when changing site or scheduler settings.