chore(skills): test NVSkills auto request

Signed-off-by: Chen Cui <chcui@nvidia.com>

Author: cuichenx

skills/nemo-mbridge-perf-cuda-graphs/SKILL.md

@@ -7,7 +7,7 @@ when_to_use: Reducing host-driver overhead via CUDA graphs, or tracing a crash o
 
 # CUDA Graphs
 
-Stable docs: @docs/training/cuda-graphs.md
+Stable documentation: @docs/training/cuda-graphs.md
 Card: @skills/nemo-mbridge-perf-cuda-graphs/card.yaml
 
 ## What It Is

skills/nemo-mbridge-perf-cuda-graphs/evals/evals.json

@@ -1 +1,42 @@
-[]
+[
+  {
+    "id": "cuda-graphs-te-scoped-moe-smoke",
+    "question": "I am training a Megatron Bridge Qwen3 MoE model and want to reduce CPU launch overhead with CUDA graphs. Which cuda_graph_impl and cuda_graph_scope should I start with, and what prerequisites should I set?",
+    "expected_skill": "nemo-mbridge-perf-cuda-graphs",
+    "expected_script": null,
+    "ground_truth": "For most Megatron Bridge training workloads, start with Transformer Engine scoped graphs rather than local full-iteration capture. For a dropless MoE model, use cuda_graph_impl=\"transformer_engine\" with cuda_graph_scope including attn, moe_router, and moe_preprocess. Set cuda_graph_warmup_steps, enable model.use_te_rng_tracker and rng.te_rng_tracker, keep sequence length and micro-batch size static, and compare steady-state replay iterations after warmup and capture. Do not combine moe and moe_router scopes.",
+    "expected_behavior": [
+      "Read the nemo-mbridge-perf-cuda-graphs skill before answering.",
+      "Recommend transformer_engine scoped graphs for the MoE bring-up path.",
+      "Name the relevant MoE scopes: attn, moe_router, and moe_preprocess.",
+      "Mention the TE RNG tracker requirement and static-shape constraint.",
+      "Tell the user to compare replay timing after warmup and capture rather than measuring the capture step."
+    ]
+  },
+  {
+    "id": "cuda-graphs-local-full-iteration-constraints",
+    "question": "My Megatron Bridge run uses cuda_graph_impl=\"local\" with cuda_graph_scope=[\"full_iteration\"], but it fails around runtime checks. What configuration constraints should I verify?",
+    "expected_skill": "nemo-mbridge-perf-cuda-graphs",
+    "expected_script": null,
+    "ground_truth": "Local full-iteration CUDA graphs are only valid with the full_iteration scope and have tighter constraints than TE scoped graphs. Verify rerun_state_machine.check_for_nan_in_loss is false, enable model.use_te_rng_tracker and rng.te_rng_tracker, keep tensor shapes static, and avoid incompatible features such as CPU offloading. If PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True is set, NCCL_GRAPH_REGISTER may need to be set to 0. The user should first stabilize the eager run and then enable the narrowest useful graph path.",
+    "expected_behavior": [
+      "Read the nemo-mbridge-perf-cuda-graphs skill before answering.",
+      "Identify that local full-iteration graphs require the full_iteration scope.",
+      "Mention that check_for_nan_in_loss must be disabled for local full_iteration graphs.",
+      "Mention TE RNG tracker and static tensor shape requirements.",
+      "Call out at least one incompatible or environment-sensitive setting such as CPU offloading or NCCL_GRAPH_REGISTER."
+    ]
+  },
+  {
+    "id": "cuda-graphs-negative-model-support",
+    "question": "How do I add a new Hugging Face model family to Megatron Bridge and wire its config mapping?",
+    "expected_skill": null,
+    "expected_script": null,
+    "ground_truth": "This prompt is about adding model support and configuration mapping, not validating or configuring CUDA graph capture. The CUDA graphs skill should not be selected.",
+    "expected_behavior": [
+      "Do not select the nemo-mbridge-perf-cuda-graphs skill.",
+      "Do not answer with CUDA graph implementation or scope guidance.",
+      "Route the task to model-support guidance or ask for the relevant model-support context instead."
+    ]
+  }
+]