Evaluate saved pruned text and VLM checkpoints

examples/puzzletron/README.md

@@ -145,7 +145,7 @@ dependency-ordered campaign and whole-campaign resume.
 Candidate evaluation can be part of a post-MIP campaign flow, where metrics,
 selection, materialization, and report lineage remain connected. Configure
 that route with the
-[post-MIP pipeline guide](docs/post_mip_pipeline.md#downstream-evaluation).
+[post-MIP pipeline guide](docs/post_mip_pipeline.md#evaluate-saved-checkpoints).
 
 For standalone evaluation in the Puzzletron worker environment, choose the
 route that matches the checkpoint and task:

examples/puzzletron/configs/families/qwen3_5/qwen3p5_0p8b/runs/full_smoke.yaml

@@ -44,16 +44,39 @@ post_mip:
         materialized:
           type: materialize
           input: best_lm
+        checkpoint_eval:
+          type: downstream_evaluation
+          input: materialized
+          failure_policy: strict
+          config: &ifeval_smoke_config
+            tasks: [ifeval]
+            limit: 2
+            batch_size: 1
+            timeout_seconds: 600
+            reasoning_parser: qwen3
+            dtype: bfloat16
+            gpu_memory_utilization: 0.5
+            max_model_len: 4096
+            topology:
+              tensor_parallel_size: 1
+              pipeline_parallel_size: 1
+              data_parallel_size: 1
+              prefill_context_parallel_size: 1
+              decode_context_parallel_size: 1
+              enable_expert_parallel: false
+              distributed_executor_backend: mp
+              gpu_group_size: 1
         serving:
           type: aiperf
-          input: materialized
+          input: checkpoint_eval
           config:
             input_tokens: 128
             output_tokens: 32
             concurrency: [1]
             request_count: 4
             use_server_token_count: true
-            benchmark_timeout: 900
+            readiness_timeout: 300
+            benchmark_timeout: 300
             allow_aiperf_v011_online_tokenizer_resolution: true
             topology:
               tensor_parallel_size: 1
@@ -84,9 +107,14 @@ post_mip:
             global_batch_size: 1
             local_batch_size: 1
             checkpoint_every_steps: 2
+        post_kd_checkpoint_eval:
+          type: downstream_evaluation
+          input: short_kd
+          failure_policy: strict
+          config: *ifeval_smoke_config
         final_eval:
           type: evaluation
-          input: short_kd
+          input: post_kd_checkpoint_eval
           config:
             eval_samples: 2
             block_size: 512

examples/puzzletron/configs/families/qwen3_5/qwen3p5_0p8b/runs/full_vlm_smoke.yaml

@@ -51,9 +51,30 @@ post_mip:
         materialized:
           type: materialize
           input: best_vlm_loss
+        checkpoint_eval:
+          type: downstream_evaluation
+          input: materialized
+          failure_policy: strict
+          config: &realworldqa_smoke_config
+            profile: qwen35_vlm_realworldqa
+            batch_size: 1
+            timeout_seconds: 600
+            dtype: bfloat16
+            gpu_memory_utilization: 0.5
+            max_model_len: 4096
+            limit_mm_per_prompt: {image: 1}
+            topology:
+              tensor_parallel_size: 1
+              pipeline_parallel_size: 1
+              data_parallel_size: 1
+              prefill_context_parallel_size: 1
+              decode_context_parallel_size: 1
+              enable_expert_parallel: false
+              distributed_executor_backend: mp
+              gpu_group_size: 1
         vlm_serving:
           type: aiperf
-          input: materialized
+          input: checkpoint_eval
           config:
             endpoint_type: chat
             input_tokens: 100
@@ -65,7 +86,8 @@ post_mip:
             request_count: 1
             extra_inputs: {min_tokens: 80}
             use_server_token_count: true
-            benchmark_timeout: 900
+            readiness_timeout: 300
+            benchmark_timeout: 300
             topology:
               tensor_parallel_size: 1
               pipeline_parallel_size: 1
@@ -100,9 +122,14 @@ post_mip:
             global_batch_size: 1
             local_batch_size: 1
             checkpoint_every_steps: 2
+        post_kd_checkpoint_eval:
+          type: downstream_evaluation
+          input: short_vlm_kd
+          failure_policy: strict
+          config: *realworldqa_smoke_config
         final_image_eval:
           type: evaluation
-          input: short_vlm_kd
+          input: post_kd_checkpoint_eval
           config:
             eval_samples: 2
             block_size: 512

examples/puzzletron/configs/orchestration/qwen3p5_0p8b/execution.full_smoke.yaml

@@ -23,8 +23,10 @@ execution:
     post.params-90.online_eval: {strategy: sharded, instances: 1}
     post.params-90.best_lm: {strategy: single, instances: 1}
     post.params-90.materialized: {strategy: sharded, instances: 1}
+    post.params-90.checkpoint_eval: {strategy: sharded, instances: 1}
     post.params-90.serving: {strategy: sharded, instances: 1}
     post.params-90.fastest: {strategy: single, instances: 1}
     post.params-90.short_kd: {strategy: sharded, instances: 1}
+    post.params-90.post_kd_checkpoint_eval: {strategy: sharded, instances: 1}
     post.params-90.final_eval: {strategy: sharded, instances: 1}
     post.params-90.best: {strategy: single, instances: 1}

examples/puzzletron/configs/orchestration/qwen3p5_0p8b/execution.full_vlm_smoke.yaml

@@ -21,8 +21,10 @@ execution:
     post.params-90.image_eval: {strategy: sharded, instances: 1}
     post.params-90.best_vlm_loss: {strategy: single, instances: 1}
     post.params-90.materialized: {strategy: sharded, instances: 1}
+    post.params-90.checkpoint_eval: {strategy: sharded, instances: 1}
     post.params-90.vlm_serving: {strategy: sharded, instances: 1}
     post.params-90.fastest_vlm: {strategy: single, instances: 1}
     post.params-90.short_vlm_kd: {strategy: sharded, instances: 1}
+    post.params-90.post_kd_checkpoint_eval: {strategy: sharded, instances: 1}
     post.params-90.final_image_eval: {strategy: sharded, instances: 1}
     post.params-90.best: {strategy: single, instances: 1}

examples/puzzletron/docs/checkpoint_evaluation.md

@@ -117,4 +117,4 @@ identity when applicable. For `lmms-eval`, retain `command.json`,
 generated YAML and launcher result artifacts.
 
 To evaluate candidates as part of a pruning campaign, use
-[downstream evaluation](post_mip_pipeline.md#downstream-evaluation) instead.
+[saved-checkpoint evaluation](post_mip_pipeline.md#evaluate-saved-checkpoints) instead.

examples/puzzletron/docs/post_mip_pipeline.md

@@ -178,15 +178,34 @@ Selection still follows `input`; `model_source` only chooses the artifact operat
 on. This supports a long KD run selected using short-KD/PTQ results but restarted
 from the original candidate.
 
-## Downstream evaluation
+## Evaluate saved checkpoints
 
 `downstream_evaluation` adapts the generic
-[checkpoint evaluator](checkpoint_evaluation.md) to materialized campaign
-candidates and publishes their task metrics. Add it after a `materialize` node;
-the linked example config shows the complete flow. Use the standalone
+[checkpoint evaluator](checkpoint_evaluation.md) to campaign checkpoints and
+publishes their task metrics. Add it after any checkpoint-producing node, such
+as `materialize` or `global_kd`. Use the standalone
 [checkpoint evaluation](checkpoint_evaluation.md) command when campaign
 lineage, filtering, and reports are not needed.
 
+Materialization writes a reloadable Hugging Face checkpoint directory. Its
+configuration records the realized per-layer block sizes, and its safetensors
+contain the physically sliced weights. Evaluation passes that saved directory
+unchanged to a fresh `lmms-eval` process backed by vLLM. It does not convert the
+AnyModel instance back into an AutoModel instance. Modality-specific profiles
+may prepare pinned task adapters and offline dataset snapshots first, but they
+delegate checkpoint execution and completion validation to the same evaluator.
+
+Global KD publishes a consolidated Hugging Face checkpoint and preserves the
+realized pruning configuration and required tokenizer or processor assets. A
+downstream evaluation node after KD therefore uses the same checkpoint contract
+as one after materialization. The Qwen 3.5 text and VLM smoke flows evaluate the
+selected checkpoint both before and after their short KD stage.
+
+The post-MIP graph does not treat the teacher as a candidate revision. To compare
+a distilled student with its teacher, evaluate the teacher separately with the
+same task, evaluator version, dataset revision, prompt settings, and sample
+limit, then compare those metrics with the post-KD node's metrics.
+
 ## Filters
 
 `top_k` accepts one integer or separate homogeneous/heterogeneous quotas.

examples/puzzletron/docs/qwen3p5_0p8b_smoke.md

@@ -3,11 +3,15 @@
 The checked-in `full_smoke` recipe runs a small end-to-end test of text-only
 pruning for Qwen 3.5 0.8B. It searches the FFN intermediate sizes
 `[3072, 2048]`, evaluates the candidates, saves the two strongest candidates as
-physical checkpoints, and measures their serving performance with AIPerf. It
-then distills the candidate with higher measured output-token throughput for
-two steps, evaluates it again, and selects the final checkpoint. The recipe
-pins the public checkpoint revision so repeated runs use the same starting
-model.
+physical checkpoints, and reloads each saved directory through vLLM for two
+IFEval samples. It then measures both checkpoints with AIPerf, distills the
+candidate with higher measured output-token throughput for two steps, evaluates
+the resulting checkpoint with another two IFEval samples and the internal
+two-sample LM-loss check, and selects the final checkpoint. The recipe pins the
+public checkpoint revision so repeated runs use the same starting model. See
+[evaluate saved checkpoints](post_mip_pipeline.md#evaluate-saved-checkpoints)
+for how both Hugging Face directories are loaded without an AnyModel-to-AutoModel
+conversion.
 
 These small budgets check that the complete workflow runs and resumes
 correctly. They do not establish model quality or production throughput.
@@ -61,9 +65,16 @@ python examples/puzzletron/orchestrate.py \
   --stage full
 ```
 
-The checked-in flow deliberately uses two evaluation samples per candidate,
-four AIPerf requests per serving candidate, and two distillation steps. These
-budgets validate lifecycle correctness, comparative serving selection, and
-resumability; they are not quality or throughput claims. Final acceptance must
-reload the selected checkpoint, verify the cumulative report, and confirm that
-the resume submits no work for completed stages.
+The checked-in flow deliberately uses two candidate-evaluation samples, two
+IFEval samples, four AIPerf requests per serving candidate, and two
+distillation steps. These budgets validate lifecycle correctness, comparative
+serving selection, and resumability; they are not quality or throughput claims.
+The worker environment must provide the [pinned evaluator
+installation](checkpoint_evaluation.md#quick-start). IFEval task data must be
+fetchable from each worker or already present in its Hugging Face cache.
+
+After completion, inspect the `checkpoint_eval` and `post_kd_checkpoint_eval`
+nodes under `artifacts/post_mip/nodes`. Their summaries must name the corresponding
+pre-KD and post-KD checkpoints, report two effective IFEval samples, and contain
+finite metrics. Also verify the cumulative report and confirm that resuming
+submits no work for completed stages.

examples/puzzletron/docs/qwen3p5_0p8b_vlm_smoke.md

@@ -4,11 +4,19 @@ The checked-in `full_vlm_smoke` recipe runs a small end-to-end test of
 vision-language pruning for the public `Qwen/Qwen3.5-0.8B` checkpoint. It uses
 real image-conversation examples to search the FFN intermediate sizes
 `[3072, 2048]`, evaluate the candidates, and save the two strongest candidates
-as physical checkpoints. It measures both checkpoints with 1-, 6-, and
-12-image AIPerf requests, distills the candidate with the highest measured
-12-image throughput for two steps, and runs a final image-and-text evaluation.
-The immutable revision in `model.yaml` ensures that repeated runs use the same
-starting model.
+as physical checkpoints. It reloads each saved directory through vLLM for two
+RealWorldQA image samples using the pinned Qwen 3.5 VLM evaluation profile. The
+profile verifies the evaluator revision and immutable offline dataset snapshot,
+strips inherited Hub credentials, and records preflight provenance before
+delegating execution to the shared checkpoint evaluator. The workflow then
+measures both checkpoints with 1-, 6-, and 12-image AIPerf requests, distills
+the candidate with the highest measured 12-image throughput for two steps, and
+runs the pinned RealWorldQA benchmark and internal image-and-text evaluation on
+the resulting checkpoint. The immutable revision in `model.yaml` ensures that
+repeated runs use the same starting model. See
+[evaluate saved checkpoints](post_mip_pipeline.md#evaluate-saved-checkpoints)
+for the direct pre-KD and post-KD reload paths; no AnyModel-to-AutoModel
+conversion occurs.
 
 These small budgets check that the complete workflow runs and resumes
 correctly. They do not establish model quality or production throughput.
@@ -18,7 +26,11 @@ correctly. They do not establish model quality or production throughput.
 Prepare the setup and worker environments described in
 [environment setup](environment_setup.md). The worker environment must provide
 ModelOpt, NeMo AutoModel's Qwen 3.5 VLM support, the Puzzletron requirements,
-and the reviewed AIPerf/vLLM runtime selected by your runner contract.
+the pinned `lmms-eval` dependency, and the reviewed AIPerf/vLLM runtime selected
+by your runner contract. The worker-visible Hugging Face cache must already
+contain the pinned RealWorldQA snapshot. Populate it as described in
+[cache benchmark data](vlm_checkpoint_evaluation.md#cache-benchmark-data); the
+profile verifies the local snapshot and evaluates it offline.
 
 The worker-visible Hugging Face cache must contain, or be allowed to fetch,
 `Qwen/Qwen3.5-0.8B` at the pinned revision in
@@ -128,11 +140,11 @@ python examples/puzzletron/orchestrate.py \
 
 Confirm that all enabled model stages use one GPU, image-backed stages resolve
 `data.modality=multimodal`, and no text tokenization stage is present. Confirm
-that two quality candidates reach `post.params-90.vlm_serving`, which declares
-a `chat` workload with 1, 6, and 12 1280x720 images per request rather than a
-text-only serving proxy. The `fastest_vlm` filter selects one candidate from
-the 12-image throughput metric before KD. Then launch by omitting only
-`--dry-run`:
+that two quality candidates reach `post.params-90.checkpoint_eval` for the
+bounded RealWorldQA run and then `post.params-90.vlm_serving`, which declares a
+`chat` workload with 1, 6, and 12 1280x720 images per request rather than a
+text-only serving proxy. The `fastest_vlm` filter selects one candidate from the
+12-image throughput metric before KD. Then launch by omitting only `--dry-run`:
 
 ```bash
 python examples/puzzletron/orchestrate.py \
@@ -157,7 +169,14 @@ canonical stage summaries:
 - width scoring and VLM KD processed real image tensors and report a nonzero
   vision-forward count;
 - sorting and physical slicing equivalence passed at the configured tolerance;
-- the selected checkpoint reloads after materialization and after VLM KD;
+- each materialized pre-KD checkpoint has a successful `checkpoint_eval`
+  summary whose `checkpoint` field names that saved artifact, whose
+  RealWorldQA sample count equals two, and whose metrics are finite; the
+  evaluation root's `profile.json` records the pinned dataset revision and
+  offline preflight;
+- the selected post-KD checkpoint has a successful `post_kd_checkpoint_eval`
+  summary for two RealWorldQA samples with finite metrics, and the internal
+  final image evaluation reloads the same checkpoint;
 - the two-step KD summary contains finite main CE/KD and MTP CE/KD metrics plus
   nonzero trainable-group gradient evidence;
 - AIPerf completes every 1-, 6-, and 12-image chat workload for both retained
@@ -195,3 +214,14 @@ vision path, and comparative selection while keeping the example bounded; it
 does not isolate vision-encoder latency or establish production performance.
 Increase request count and concurrency in a separate reviewed performance run
 before making throughput claims.
+
+The real-checkpoint lifecycle test is opt-in and is not part of default pytest
+or routine CI smoke execution. Point it at the populated cache:
+
+```bash
+PUZZLETRON_VLM_BENCHMARK_HF_HOME=/path/to/hf-home \
+  pytest --run-manual tests/gpu/torch/puzzletron/test_qwen3p5_0p8b_vlm_smoke.py
+```
+
+The checkpoint contract is documented in
+[evaluate saved checkpoints](post_mip_pipeline.md#evaluate-saved-checkpoints).

examples/puzzletron/evaluation/vlm/__init__.py

@@ -0,0 +1,18 @@
+# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""Public interfaces for Puzzletron VLM evaluation."""
+
+from .evaluator import *