From cdb9c50daa75cef8ded7533a5b5b642af3fe8bfd Mon Sep 17 00:00:00 2001 From: Chen Cui Date: Thu, 28 May 2026 15:59:14 -0700 Subject: [PATCH 1/6] [skills] rename overlap and recompute skills with nemo mbridge prefix Signed-off-by: Chen Cui --- docs/skills-index.md | 6 +++--- .../SKILL.md | 9 +++++---- .../card.yaml | 2 +- .../evals/evals.json | 1 + .../SKILL.md | 7 ++++--- .../card.yaml | 0 .../evals/evals.json | 1 + .../SKILL.md | 5 +++-- .../card.yaml | 0 .../nemo-mbridge-perf-moe-comm-overlap/evals/evals.json | 1 + 10 files changed, 19 insertions(+), 13 deletions(-) rename skills/{perf-activation-recompute => nemo-mbridge-perf-activation-recompute}/SKILL.md (96%) rename skills/{perf-activation-recompute => nemo-mbridge-perf-activation-recompute}/card.yaml (98%) create mode 100644 skills/nemo-mbridge-perf-activation-recompute/evals/evals.json rename skills/{perf-expert-parallel-overlap => nemo-mbridge-perf-expert-parallel-overlap}/SKILL.md (98%) rename skills/{perf-expert-parallel-overlap => nemo-mbridge-perf-expert-parallel-overlap}/card.yaml (100%) create mode 100644 skills/nemo-mbridge-perf-expert-parallel-overlap/evals/evals.json rename skills/{perf-moe-comm-overlap => nemo-mbridge-perf-moe-comm-overlap}/SKILL.md (97%) rename skills/{perf-moe-comm-overlap => nemo-mbridge-perf-moe-comm-overlap}/card.yaml (100%) create mode 100644 skills/nemo-mbridge-perf-moe-comm-overlap/evals/evals.json diff --git a/docs/skills-index.md b/docs/skills-index.md index 7a2efbf3f0..58d3d6f3dc 100644 --- a/docs/skills-index.md +++ b/docs/skills-index.md @@ -40,9 +40,9 @@ skills/perf-tp-dp-comm-overlap/SKILL skills/perf-megatron-fsdp/SKILL skills/perf-sequence-packing/SKILL skills/perf-hierarchical-context-parallel/SKILL -skills/perf-expert-parallel-overlap/SKILL -skills/perf-moe-comm-overlap/SKILL -skills/perf-activation-recompute/SKILL +skills/nemo-mbridge-perf-expert-parallel-overlap/SKILL +skills/nemo-mbridge-perf-moe-comm-overlap/SKILL +skills/nemo-mbridge-perf-activation-recompute/SKILL skills/perf-memory-tuning/SKILL skills/perf-moe-dispatcher-selection/SKILL skills/perf-moe-hardware-configs/SKILL diff --git a/skills/perf-activation-recompute/SKILL.md b/skills/nemo-mbridge-perf-activation-recompute/SKILL.md similarity index 96% rename from skills/perf-activation-recompute/SKILL.md rename to skills/nemo-mbridge-perf-activation-recompute/SKILL.md index cbb574327d..3eb1570ba9 100644 --- a/skills/perf-activation-recompute/SKILL.md +++ b/skills/nemo-mbridge-perf-activation-recompute/SKILL.md @@ -1,13 +1,14 @@ --- -name: perf-activation-recompute +name: nemo-mbridge-perf-activation-recompute description: Validate and use selective and full activation recompute in Megatron Bridge to reduce GPU memory usage at the cost of extra compute. +license: Apache-2.0 when_to_use: Reducing GPU memory via activation recompute, or investigating a commit that changed recompute settings and caused OOM or a regression; 'recompute_granularity', 'recompute_num_layers', 'recompute_modules', 'recompute_method', 'selective recompute', 'full recompute', 'activation memory OOM'. --- # Activation Recompute Stable docs: @docs/training/activation-recomputation.md -Card: @skills/perf-activation-recompute/card.yaml +Card: @skills/nemo-mbridge-perf-activation-recompute/card.yaml ## What It Is @@ -28,7 +29,7 @@ how many layers via `recompute_num_layers`. 1. **Set `PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True` first** — most borderline OOMs are caused by memory fragmentation, not capacity. This - fixes it at zero cost. See @skills/perf-memory-tuning/SKILL.md. + fixes it at zero cost. See @skills/mbridge-perf-memory-tuning/SKILL.md. 2. Start with `recompute_granularity=selective`, `recompute_modules=[core_attn]` (often already the default in recipes). 3. Add `layernorm` to recompute modules — nearly free compute-wise but saves @@ -128,7 +129,7 @@ Key takeaways: (hidden=28672) is expensive to recompute - Combining `mlp` + `core_attn` is slightly worse than `mlp` alone - For this workload, the actual OOM fix was `PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True` - (memory fragmentation, not capacity). See @skills/perf-memory-tuning/SKILL.md. + (memory fragmentation, not capacity). See @skills/mbridge-perf-memory-tuning/SKILL.md. ## Code Anchors diff --git a/skills/perf-activation-recompute/card.yaml b/skills/nemo-mbridge-perf-activation-recompute/card.yaml similarity index 98% rename from skills/perf-activation-recompute/card.yaml rename to skills/nemo-mbridge-perf-activation-recompute/card.yaml index 96f6d8ea36..2f2a1a5832 100644 --- a/skills/perf-activation-recompute/card.yaml +++ b/skills/nemo-mbridge-perf-activation-recompute/card.yaml @@ -92,7 +92,7 @@ recommended_path: first_try: "recompute_granularity=selective, recompute_modules=[core_attn]" if_still_oom: "add layernorm (cheap) or mlp (expensive but saves ~3 GB)" last_resort: "recompute_granularity=full, recompute_method=uniform" - alternative: "see skills/perf-memory-tuning/ for VPP tuning and other memory strategies" + alternative: "see skills/mbridge-perf-memory-tuning/ for VPP tuning and other memory strategies" expected_metric_change: - metric: peak_memory direction: down diff --git a/skills/nemo-mbridge-perf-activation-recompute/evals/evals.json b/skills/nemo-mbridge-perf-activation-recompute/evals/evals.json new file mode 100644 index 0000000000..fe51488c70 --- /dev/null +++ b/skills/nemo-mbridge-perf-activation-recompute/evals/evals.json @@ -0,0 +1 @@ +[] diff --git a/skills/perf-expert-parallel-overlap/SKILL.md b/skills/nemo-mbridge-perf-expert-parallel-overlap/SKILL.md similarity index 98% rename from skills/perf-expert-parallel-overlap/SKILL.md rename to skills/nemo-mbridge-perf-expert-parallel-overlap/SKILL.md index f9164349a5..4b712132f4 100644 --- a/skills/perf-expert-parallel-overlap/SKILL.md +++ b/skills/nemo-mbridge-perf-expert-parallel-overlap/SKILL.md @@ -1,18 +1,19 @@ --- -name: perf-expert-parallel-overlap +name: nemo-mbridge-perf-expert-parallel-overlap description: Validate and use MoE expert-parallel communication overlap in Megatron-Bridge, including overlap_moe_expert_parallel_comm, delay_wgrad_compute, and flex dispatcher backends such as DeepEP and HybridEP. +license: Apache-2.0 when_to_use: Enabling EP overlap to hide dispatch/combine latency, or tracing a throughput regression to an EP overlap config change; 'overlap_moe_expert_parallel_comm', 'delay_wgrad_compute', 'flex dispatcher', 'DeepEP overlap', 'HybridEP overlap'. --- # MoE Expert-Parallel Overlap Skill Stable docs: @docs/training/communication-overlap.md -Card: @skills/perf-expert-parallel-overlap/card.yaml +Card: @skills/nemo-mbridge-perf-expert-parallel-overlap/card.yaml ## References - Stable docs: @docs/training/communication-overlap.md -- Structured metadata: @skills/perf-expert-parallel-overlap/card.yaml +- Structured metadata: @skills/nemo-mbridge-perf-expert-parallel-overlap/card.yaml ## What It Is diff --git a/skills/perf-expert-parallel-overlap/card.yaml b/skills/nemo-mbridge-perf-expert-parallel-overlap/card.yaml similarity index 100% rename from skills/perf-expert-parallel-overlap/card.yaml rename to skills/nemo-mbridge-perf-expert-parallel-overlap/card.yaml diff --git a/skills/nemo-mbridge-perf-expert-parallel-overlap/evals/evals.json b/skills/nemo-mbridge-perf-expert-parallel-overlap/evals/evals.json new file mode 100644 index 0000000000..fe51488c70 --- /dev/null +++ b/skills/nemo-mbridge-perf-expert-parallel-overlap/evals/evals.json @@ -0,0 +1 @@ +[] diff --git a/skills/perf-moe-comm-overlap/SKILL.md b/skills/nemo-mbridge-perf-moe-comm-overlap/SKILL.md similarity index 97% rename from skills/perf-moe-comm-overlap/SKILL.md rename to skills/nemo-mbridge-perf-moe-comm-overlap/SKILL.md index 7612526eeb..be69bef379 100644 --- a/skills/perf-moe-comm-overlap/SKILL.md +++ b/skills/nemo-mbridge-perf-moe-comm-overlap/SKILL.md @@ -1,6 +1,7 @@ --- -name: perf-moe-comm-overlap +name: nemo-mbridge-perf-moe-comm-overlap description: MoE expert-parallel communication overlap in Megatron Bridge. Covers dispatch/combine overlap, flex dispatcher backends, and expert wgrad scheduling. +license: Apache-2.0 when_to_use: Tuning MoE communication overlap, or tracing a MoE throughput regression to a comm-overlap config change; 'overlap_moe_expert_parallel_comm', 'MoE dispatch overlap', 'flex dispatcher', 'DeepEP overlap', 'expert wgrad scheduling'. --- @@ -9,7 +10,7 @@ when_to_use: Tuning MoE communication overlap, or tracing a MoE throughput regre For the higher-level overview, see: - @docs/training/communication-overlap.md -- @skills/perf-moe-comm-overlap/card.yaml +- @skills/nemo-mbridge-perf-moe-comm-overlap/card.yaml ## Quick Decision diff --git a/skills/perf-moe-comm-overlap/card.yaml b/skills/nemo-mbridge-perf-moe-comm-overlap/card.yaml similarity index 100% rename from skills/perf-moe-comm-overlap/card.yaml rename to skills/nemo-mbridge-perf-moe-comm-overlap/card.yaml diff --git a/skills/nemo-mbridge-perf-moe-comm-overlap/evals/evals.json b/skills/nemo-mbridge-perf-moe-comm-overlap/evals/evals.json new file mode 100644 index 0000000000..fe51488c70 --- /dev/null +++ b/skills/nemo-mbridge-perf-moe-comm-overlap/evals/evals.json @@ -0,0 +1 @@ +[] From d99e20b446767a9f028903a8380eb108425a11cf Mon Sep 17 00:00:00 2001 From: nvskills-svc-account Date: Thu, 28 May 2026 23:12:22 +0000 Subject: [PATCH 2/6] Attach NVSkills validation signatures Signed-off-by: nvskills-svc-account --- .../BENCHMARK.md | 65 +++++++++++++++++++ .../skill-card.md | 52 +++++++++++++++ .../skill.oms.sig | 1 + .../BENCHMARK.md | 65 +++++++++++++++++++ .../skill-card.md | 49 ++++++++++++++ .../skill.oms.sig | 1 + .../BENCHMARK.md | 64 ++++++++++++++++++ .../skill-card.md | 53 +++++++++++++++ .../skill.oms.sig | 1 + 9 files changed, 351 insertions(+) create mode 100644 skills/nemo-mbridge-perf-activation-recompute/BENCHMARK.md create mode 100644 skills/nemo-mbridge-perf-activation-recompute/skill-card.md create mode 100644 skills/nemo-mbridge-perf-activation-recompute/skill.oms.sig create mode 100644 skills/nemo-mbridge-perf-expert-parallel-overlap/BENCHMARK.md create mode 100644 skills/nemo-mbridge-perf-expert-parallel-overlap/skill-card.md create mode 100644 skills/nemo-mbridge-perf-expert-parallel-overlap/skill.oms.sig create mode 100644 skills/nemo-mbridge-perf-moe-comm-overlap/BENCHMARK.md create mode 100644 skills/nemo-mbridge-perf-moe-comm-overlap/skill-card.md create mode 100644 skills/nemo-mbridge-perf-moe-comm-overlap/skill.oms.sig diff --git a/skills/nemo-mbridge-perf-activation-recompute/BENCHMARK.md b/skills/nemo-mbridge-perf-activation-recompute/BENCHMARK.md new file mode 100644 index 0000000000..e2ae5dd1f4 --- /dev/null +++ b/skills/nemo-mbridge-perf-activation-recompute/BENCHMARK.md @@ -0,0 +1,65 @@ +# Evaluation Report + +Evaluation of the `nemo-mbridge-perf-activation-recompute` skill before publication through NVSkills-Eval. + +This benchmark summarizes 3-Tier Evaluation from NVSkills-Eval results for the skill. The goal is to document whether the skill is safe, discoverable, effective, and useful for agents before it is published for broader workflow use. + +## Evaluation Summary + +- Skill: `nemo-mbridge-perf-activation-recompute` +- Evaluation date: 2026-05-28 +- NVSkills-Eval profile: `external` +- Overall verdict: FAIL +- Tier 3 live agent evaluation: not available in this report + +## Agents Used + +- Tier 3 agent details were not available in this report. + +## Metrics Used + +Reported benchmark dimensions: + +- Security: checks whether skill-assisted execution avoids unsafe behavior such as secret leakage, destructive commands, or unauthorized access. +- Correctness: checks whether the agent follows the expected workflow and produces the correct final output. +- Discoverability: checks whether the agent loads the skill when relevant and avoids using it when irrelevant. +- Effectiveness: checks whether the agent performs measurably better with the skill than without it. +- Efficiency: checks whether the agent uses fewer tokens and avoids redundant work. + +Underlying evaluation signals used in this run: + +- No Tier 3 evaluation signal details were available in this report. + +## Test Tasks + +Tier 3 evaluation task details were not available in this report. + +## Results + +Tier 3 dimension rollup was not available in this report. + +## Tier 1: Static Validation Summary + +Tier 1 validation passed with observations. NVSkills-Eval ran 9 checks and found 10 total findings. + +Top findings: + +- MEDIUM QUALITY/quality_correctness: SKILL_SPEC recommended field missing: 'metadata.author' (`skills/nemo-mbridge-perf-activation-recompute/SKILL.md`) +- MEDIUM QUALITY/quality_correctness: SKILL_SPEC recommended field missing: 'metadata.tags' (`skills/nemo-mbridge-perf-activation-recompute/SKILL.md`) +- MEDIUM SCHEMA/body_recommended_section: Missing recommended section: '## Instructions' (`skills/nemo-mbridge-perf-activation-recompute/SKILL.md`) +- MEDIUM SCHEMA/body_recommended_section: Missing recommended section: '## Examples' (`skills/nemo-mbridge-perf-activation-recompute/SKILL.md`) +- MEDIUM SCHEMA/author_missing: Author not specified in metadata (`skills/nemo-mbridge-perf-activation-recompute/SKILL.md`) + +## Tier 2: Deduplication Summary + +Tier 2 validation reported findings. NVSkills-Eval ran 2 checks and found 1 total findings. + +Top findings: + +- HIGH DUPLICATE/duplicate: Duplicate content found within SKILL.md: + "### Selective recompute (default for most recipes)" in SKILL.md (lines 40-46) + vs "### Selective recompute with additional modules" in SKILL.md (lines 47-53) (`SKILL.md:40`) + +## Publication Recommendation + +The skill should be reviewed before NVSkills-Eval publication. Skill owners should address the findings above and rerun NVSkills-Eval to refresh this benchmark. diff --git a/skills/nemo-mbridge-perf-activation-recompute/skill-card.md b/skills/nemo-mbridge-perf-activation-recompute/skill-card.md new file mode 100644 index 0000000000..cfd22d1c7a --- /dev/null +++ b/skills/nemo-mbridge-perf-activation-recompute/skill-card.md @@ -0,0 +1,52 @@ +## Description:
+Validate and use selective and full activation recompute in Megatron Bridge to reduce GPU memory usage at the cost of extra compute.
+ +This skill is ready for commercial/non-commercial use.
+ +## Owner +NVIDIA
+ +### License/Terms of Use:
+Apache 2.0
+## Use Case:
+Developers and engineers configuring activation recompute strategies in Megatron Bridge to reduce GPU memory consumption during large language model training, or diagnosing OOM and performance regressions caused by recompute settings.
+ +### Deployment Geography for Use:
+Global
+ +## Known Risks and Mitigations:
+Risk: Review before execution as proposals could introduce incorrect or misleading guidance into skills.
+Mitigation: Review and scan skill before deployment.
+ +## Reference(s):
+- [Activation Recomputation Documentation](docs/training/activation-recomputation.md)
+- [Performance Tuning Guide](docs/performance-guide.md)
+ + +## Skill Output:
+**Output Type(s):** [Configuration instructions, Shell commands, Analysis]
+**Output Format:** [Markdown with inline Python and bash code blocks]
+**Output Parameters:** [1D]
+**Other Properties Related to Output:** [None]
+ +## Evaluation Tasks:
+NVSkills-Eval 3-Tier evaluation (external profile) assessing skill safety, discoverability, effectiveness, and usefulness for agents before publication.
+ +## Evaluation Metrics Used:
+Reported benchmark dimensions:
+- Security: Checks whether skill-assisted execution avoids unsafe behavior such as secret leakage, destructive commands, or unauthorized access.
+- Correctness: Checks whether the agent follows the expected workflow and produces the correct final output.
+- Discoverability: Checks whether the agent loads the skill when relevant and avoids using it when irrelevant.
+- Effectiveness: Checks whether the agent performs measurably better with the skill than without it.
+- Efficiency: Checks whether the agent uses fewer tokens and avoids redundant work.
+ + + +## Skill Version(s):
+v0.2.0rc6-1465-gcdb9c50d (source: git describe)
+ +## Ethical Considerations:
+NVIDIA believes Trustworthy AI is a shared responsibility and we have established policies and practices to enable development for a wide array of AI applications. When downloaded or used in accordance with our terms of service, developers should work with their internal team to ensure this skill meets requirements for the relevant industry and use case and addresses unforeseen product misuse.
+ +(For Release on NVIDIA Platforms Only)
+Please report quality, risk, security vulnerabilities or NVIDIA AI Concerns [here](https://app.intigriti.com/programs/nvidia/nvidiavdp/detail).
diff --git a/skills/nemo-mbridge-perf-activation-recompute/skill.oms.sig b/skills/nemo-mbridge-perf-activation-recompute/skill.oms.sig new file mode 100644 index 0000000000..72804f953b --- /dev/null +++ b/skills/nemo-mbridge-perf-activation-recompute/skill.oms.sig @@ -0,0 +1 @@ +{"mediaType":"application/vnd.dev.sigstore.bundle.v0.3+json","verificationMaterial":{"x509CertificateChain":{"certificates":[{"rawBytes":"MIICgzCCAgmgAwIBAgIUKIyS7SxNteQIiWzK1dWj85E6520wCgYIKoZIzj0EAwMwVTELMAkGA1UEBhMCVVMxGzAZBgNVBAoMEk5WSURJQSBDb3Jwb3JhdGlvbjEpMCcGA1UEAwwgTlZJRElBIEFnZW50IENhcGFiaWxpdGllcyBJQ0EgMDEwHhcNMjYwNDAxMDAwMDAwWhcNMjgwNDIyMTUzMzA5WjBUMQswCQYDVQQGEwJVUzEbMBkGA1UECgwSTlZJRElBIENvcnBvcmF0aW9uMSgwJgYDVQQDDB9OVklESUEgQWdlbnQgU2tpbGxzIFNpZ25pbmcgMDAxMHYwEAYHKoZIzj0CAQYFK4EEACIDYgAEYoRM9bQl/dGlwSRNi6bTpIJUXH8Nv9GciP6LSflJYYMLCc296kpyuTSsk5ddbAWiDcFX3C/ydX3jwc+qCLYP6uHy9XphyLjOQ27Yb2J6rBLVtRBS1mgGco/Gr7fL6ODco4GaMIGXMB0GA1UdDgQWBBRQ/5ZW3nJ6lmo9SVk7I15o7UGmpTAfBgNVHSMEGDAWgBRPGpILxMBBleJSsBGjrMKsby1CgjAMBgNVHRMBAf8EAjAAMA4GA1UdDwEB/wQEAwIHgDA3BggrBgEFBQcBAQQrMCkwJwYIKwYBBQUHMAGGG2h0dHA6Ly9vY3NwLm5kaXMubnZpZGlhLmNvbTAKBggqhkjOPQQDAwNoADBlAjAUygu/GiOCIXrgGr4SmLgeEVDcEitfFUv7ALbvLVGVyMysB3mxmO/uInZfXzWcJZsCMQDxuoxj4ZmO30jhkPIcCxGFCOvnUsnfU3TfGcouYm4M6iRpbKvtVnHPiy4bi6pcKf0="},{"rawBytes":"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"},{"rawBytes":"MIICRTCCAcygAwIBAgIUeJdY3rV86EdvFmG7L8LJBsyQFYkwCgYIKoZIzj0EAwMwUTELMAkGA1UEBhMCVVMxGzAZBgNVBAoMEk5WSURJQSBDb3Jwb3JhdGlvbjElMCMGA1UEAwwcTlZJRElBIEFnZW50IENhcGFiaWxpdGllcyBDQTAgFw0yNjA0MDEwMDAwMDBaGA85OTk5MTIzMTIzNTk1OVowUTELMAkGA1UEBhMCVVMxGzAZBgNVBAoMEk5WSURJQSBDb3Jwb3JhdGlvbjElMCMGA1UEAwwcTlZJRElBIEFnZW50IENhcGFiaWxpdGllcyBDQTB2MBAGByqGSM49AgEGBSuBBAAiA2IABAYpiXCDjJ9NT2eSDhyHJVSw1Tbze18cGG2F/578oWvHxg23eQAhNRYdq88i1iOshZSO6C29doKui5Xpmo/7Ctw9Sx4PP2RzOmIuOLCuTdNtKcTRwi4GEsd5BAFvWj42M6NjMGEwHQYDVR0OBBYEFItnoAjjfuCEUvzyvWyI2vOGvwPjMB8GA1UdIwQYMBaAFItnoAjjfuCEUvzyvWyI2vOGvwPjMA8GA1UdEwEB/wQFMAMBAf8wDgYDVR0PAQH/BAQDAgEGMAoGCCqGSM49BAMDA2cAMGQCMCwtAjWLaNwgGWNCgdyNoTyvNhqWRECRJV2r3+7w8g0PL6NHLOsbkgE09BH95h8XlgIwTaQmbbUh2ChAJ5TA1wRiVDnCcvbzHlZl2jM2FcwQQZlk19LOAbyGMRixbu2Ww/rj"}]},"tlogEntries":[]},"dsseEnvelope":{"payload":"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","payloadType":"application/vnd.in-toto+json","signatures":[{"sig":"MGUCMQD6eiJvxXq6hfdLhIrDVw+3soHxglRo8sb2LIjjuFp7LFrDtBaH04LNwR+2y40Jfm4CMEW6ZQZ3//bMyVqzB1V8q1EW+MOwEhjlsZh4S6QIZpJCP42j//lLgsegHxwBAhmgXw==","keyid":""}]}} \ No newline at end of file diff --git a/skills/nemo-mbridge-perf-expert-parallel-overlap/BENCHMARK.md b/skills/nemo-mbridge-perf-expert-parallel-overlap/BENCHMARK.md new file mode 100644 index 0000000000..e0f009499d --- /dev/null +++ b/skills/nemo-mbridge-perf-expert-parallel-overlap/BENCHMARK.md @@ -0,0 +1,65 @@ +# Evaluation Report + +Evaluation of the `nemo-mbridge-perf-expert-parallel-overlap` skill before publication through NVSkills-Eval. + +This benchmark summarizes 3-Tier Evaluation from NVSkills-Eval results for the skill. The goal is to document whether the skill is safe, discoverable, effective, and useful for agents before it is published for broader workflow use. + +## Evaluation Summary + +- Skill: `nemo-mbridge-perf-expert-parallel-overlap` +- Evaluation date: 2026-05-28 +- NVSkills-Eval profile: `external` +- Overall verdict: FAIL +- Tier 3 live agent evaluation: not available in this report + +## Agents Used + +- Tier 3 agent details were not available in this report. + +## Metrics Used + +Reported benchmark dimensions: + +- Security: checks whether skill-assisted execution avoids unsafe behavior such as secret leakage, destructive commands, or unauthorized access. +- Correctness: checks whether the agent follows the expected workflow and produces the correct final output. +- Discoverability: checks whether the agent loads the skill when relevant and avoids using it when irrelevant. +- Effectiveness: checks whether the agent performs measurably better with the skill than without it. +- Efficiency: checks whether the agent uses fewer tokens and avoids redundant work. + +Underlying evaluation signals used in this run: + +- No Tier 3 evaluation signal details were available in this report. + +## Test Tasks + +Tier 3 evaluation task details were not available in this report. + +## Results + +Tier 3 dimension rollup was not available in this report. + +## Tier 1: Static Validation Summary + +Tier 1 validation passed with observations. NVSkills-Eval ran 9 checks and found 11 total findings. + +Top findings: + +- MEDIUM QUALITY/quality_correctness: SKILL_SPEC recommended field missing: 'metadata.author' (`skills/nemo-mbridge-perf-expert-parallel-overlap/SKILL.md`) +- MEDIUM QUALITY/quality_correctness: SKILL_SPEC recommended field missing: 'metadata.tags' (`skills/nemo-mbridge-perf-expert-parallel-overlap/SKILL.md`) +- MEDIUM SCHEMA/body_recommended_section: Missing recommended section: '## Instructions' (`skills/nemo-mbridge-perf-expert-parallel-overlap/SKILL.md`) +- MEDIUM SCHEMA/body_recommended_section: Missing recommended section: '## Examples' (`skills/nemo-mbridge-perf-expert-parallel-overlap/SKILL.md`) +- MEDIUM SCHEMA/author_missing: Author not specified in metadata (`skills/nemo-mbridge-perf-expert-parallel-overlap/SKILL.md`) + +## Tier 2: Deduplication Summary + +Tier 2 validation reported findings. NVSkills-Eval ran 2 checks and found 1 total findings. + +Top findings: + +- HIGH DUPLICATE/duplicate: Duplicate content found within SKILL.md: + "# MoE Expert-Parallel Overlap Skill" in SKILL.md (lines 1-5) + vs "## References" in SKILL.md (lines 6-10) (`SKILL.md:1`) + +## Publication Recommendation + +The skill should be reviewed before NVSkills-Eval publication. Skill owners should address the findings above and rerun NVSkills-Eval to refresh this benchmark. diff --git a/skills/nemo-mbridge-perf-expert-parallel-overlap/skill-card.md b/skills/nemo-mbridge-perf-expert-parallel-overlap/skill-card.md new file mode 100644 index 0000000000..b0096b6df1 --- /dev/null +++ b/skills/nemo-mbridge-perf-expert-parallel-overlap/skill-card.md @@ -0,0 +1,49 @@ +## Description:
+Validate and use MoE expert-parallel communication overlap in Megatron-Bridge, including overlap_moe_expert_parallel_comm, delay_wgrad_compute, and flex dispatcher backends such as DeepEP and HybridEP.
+ +This skill is ready for commercial/non-commercial use.
+ +## Owner +NVIDIA
+ +### License/Terms of Use:
+Apache 2.0
+## Use Case:
+Developers and engineers enabling expert-parallel communication overlap to hide dispatch/combine latency in MoE model training, or diagnosing throughput regressions related to EP overlap configuration changes in Megatron-Bridge.
+ +### Deployment Geography for Use:
+Global
+ +## Known Risks and Mitigations:
+Risk: Review before execution as proposals could introduce incorrect or misleading guidance into skills.
+Mitigation: Review and scan skill before deployment.
+ +## Reference(s):
+- [Performance Tuning Guide](docs/performance-guide.md)
+- [Megatron-Bridge GitHub Repository](https://github.com/NVIDIA-NeMo/Megatron-Bridge)
+ + +## Skill Output:
+**Output Type(s):** [Shell commands, Configuration instructions, Analysis]
+**Output Format:** [Markdown with inline Python and bash code blocks]
+**Output Parameters:** [1D]
+**Other Properties Related to Output:** [None]
+ +## Evaluation Metrics Used:
+Reported benchmark dimensions:
+- Security: Checks whether skill-assisted execution avoids unsafe behavior such as secret leakage, destructive commands, or unauthorized access.
+- Correctness: Checks whether the agent follows the expected workflow and produces the correct final output.
+- Discoverability: Checks whether the agent loads the skill when relevant and avoids using it when irrelevant.
+- Effectiveness: Checks whether the agent performs measurably better with the skill than without it.
+- Efficiency: Checks whether the agent uses fewer tokens and avoids redundant work.
+ + + +## Skill Version(s):
+v0.2.0rc6-1465-gcdb9c50d (source: git describe)
+ +## Ethical Considerations:
+NVIDIA believes Trustworthy AI is a shared responsibility and we have established policies and practices to enable development for a wide array of AI applications. When downloaded or used in accordance with our terms of service, developers should work with their internal team to ensure this skill meets requirements for the relevant industry and use case and addresses unforeseen product misuse.
+ +(For Release on NVIDIA Platforms Only)
+Please report quality, risk, security vulnerabilities or NVIDIA AI Concerns [here](https://app.intigriti.com/programs/nvidia/nvidiavdp/detail).
diff --git a/skills/nemo-mbridge-perf-expert-parallel-overlap/skill.oms.sig b/skills/nemo-mbridge-perf-expert-parallel-overlap/skill.oms.sig new file mode 100644 index 0000000000..afaa5c29af --- /dev/null +++ b/skills/nemo-mbridge-perf-expert-parallel-overlap/skill.oms.sig @@ -0,0 +1 @@ +{"mediaType":"application/vnd.dev.sigstore.bundle.v0.3+json","verificationMaterial":{"x509CertificateChain":{"certificates":[{"rawBytes":"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"},{"rawBytes":"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"},{"rawBytes":"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"}]},"tlogEntries":[]},"dsseEnvelope":{"payload":"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","payloadType":"application/vnd.in-toto+json","signatures":[{"sig":"MGQCMEg6td5XT4+jcUKtqo0MLG/CgEVPEuMmkOVXYGvbF/37U+p76zzEKU5r9arNe1ov7gIwQW6InESVYD/M/kNfBo74eZtA/HG9F5bHGAPR88Jw+yPujbAYVjeBrebtgLANRYb7","keyid":""}]}} \ No newline at end of file diff --git a/skills/nemo-mbridge-perf-moe-comm-overlap/BENCHMARK.md b/skills/nemo-mbridge-perf-moe-comm-overlap/BENCHMARK.md new file mode 100644 index 0000000000..c08ed4b807 --- /dev/null +++ b/skills/nemo-mbridge-perf-moe-comm-overlap/BENCHMARK.md @@ -0,0 +1,64 @@ +# Evaluation Report + +Evaluation of the `nemo-mbridge-perf-moe-comm-overlap` skill before publication through NVSkills-Eval. + +This benchmark summarizes 3-Tier Evaluation from NVSkills-Eval results for the skill. The goal is to document whether the skill is safe, discoverable, effective, and useful for agents before it is published for broader workflow use. + +## Evaluation Summary + +- Skill: `nemo-mbridge-perf-moe-comm-overlap` +- Evaluation date: 2026-05-28 +- NVSkills-Eval profile: `external` +- Overall verdict: PASS +- Tier 3 live agent evaluation: not available in this report + +## Agents Used + +- Tier 3 agent details were not available in this report. + +## Metrics Used + +Reported benchmark dimensions: + +- Security: checks whether skill-assisted execution avoids unsafe behavior such as secret leakage, destructive commands, or unauthorized access. +- Correctness: checks whether the agent follows the expected workflow and produces the correct final output. +- Discoverability: checks whether the agent loads the skill when relevant and avoids using it when irrelevant. +- Effectiveness: checks whether the agent performs measurably better with the skill than without it. +- Efficiency: checks whether the agent uses fewer tokens and avoids redundant work. + +Underlying evaluation signals used in this run: + +- No Tier 3 evaluation signal details were available in this report. + +## Test Tasks + +Tier 3 evaluation task details were not available in this report. + +## Results + +Tier 3 dimension rollup was not available in this report. + +## Tier 1: Static Validation Summary + +Tier 1 validation passed with observations. NVSkills-Eval ran 9 checks and found 10 total findings. + +Top findings: + +- MEDIUM QUALITY/quality_correctness: SKILL_SPEC recommended field missing: 'metadata.author' (`skills/nemo-mbridge-perf-moe-comm-overlap/SKILL.md`) +- MEDIUM QUALITY/quality_correctness: SKILL_SPEC recommended field missing: 'metadata.tags' (`skills/nemo-mbridge-perf-moe-comm-overlap/SKILL.md`) +- MEDIUM SCHEMA/body_recommended_section: Missing recommended section: '## Instructions' (`skills/nemo-mbridge-perf-moe-comm-overlap/SKILL.md`) +- MEDIUM SCHEMA/body_recommended_section: Missing recommended section: '## Examples' (`skills/nemo-mbridge-perf-moe-comm-overlap/SKILL.md`) +- MEDIUM SCHEMA/author_missing: Author not specified in metadata (`skills/nemo-mbridge-perf-moe-comm-overlap/SKILL.md`) + +## Tier 2: Deduplication Summary + +Tier 2 validation passed. NVSkills-Eval ran 2 checks and found 0 total findings. + +Notable observations: + +- Context Deduplication: Collected 1 file(s) +- Inter-Skill Deduplication: Parsed skill 'nemo-mbridge-perf-moe-comm-overlap': 149 char description + +## Publication Recommendation + +The skill is suitable to proceed toward NVSkills-Eval publication based on this benchmark. Skill owners should keep this file with the skill and refresh it when the evaluation dataset, skill behavior, or target agents materially change. diff --git a/skills/nemo-mbridge-perf-moe-comm-overlap/skill-card.md b/skills/nemo-mbridge-perf-moe-comm-overlap/skill-card.md new file mode 100644 index 0000000000..53a23f0e6b --- /dev/null +++ b/skills/nemo-mbridge-perf-moe-comm-overlap/skill-card.md @@ -0,0 +1,53 @@ +## Description:
+MoE expert-parallel communication overlap in Megatron Bridge. Covers dispatch/combine overlap, flex dispatcher backends, and expert wgrad scheduling.
+ +This skill is ready for commercial/non-commercial use.
+ +## Owner +NVIDIA
+ +### License/Terms of Use:
+Apache 2.0
+## Use Case:
+Developers and engineers tuning MoE communication overlap for expert-parallel training in Megatron Bridge, or tracing throughput regressions to comm-overlap configuration changes.
+ +### Deployment Geography for Use:
+Global
+ +## Known Risks and Mitigations:
+Risk: Review before execution as proposals could introduce incorrect or misleading guidance into skills.
+Mitigation: Review and scan skill before deployment.
+ +## Reference(s):
+- [Communication Overlap Documentation](docs/training/communication-overlap.md)
+- [Performance Tuning Guide](docs/performance-guide.md)
+- [Parallelisms Documentation](docs/parallelisms.md)
+ + +## Skill Output:
+**Output Type(s):** [Configuration instructions, Shell commands, Analysis]
+**Output Format:** [Markdown with inline code blocks]
+**Output Parameters:** [1D]
+**Other Properties Related to Output:** [None]
+ +## Evaluation Tasks:
+NVSkills-Eval 3-Tier evaluation with external profile. Tier 1 static validation (9 checks), Tier 2 deduplication (2 checks). Tier 3 live agent evaluation not available.
+ +## Evaluation Metrics Used:
+Reported benchmark dimensions:
+- Security: Checks whether skill-assisted execution avoids unsafe behavior such as secret leakage, destructive commands, or unauthorized access.
+- Correctness: Checks whether the agent follows the expected workflow and produces the correct final output.
+- Discoverability: Checks whether the agent loads the skill when relevant and avoids using it when irrelevant.
+- Effectiveness: Checks whether the agent performs measurably better with the skill than without it.
+- Efficiency: Checks whether the agent uses fewer tokens and avoids redundant work.
+ + + +## Skill Version(s):
+cdb9c50d (source: git SHA, committed 2026-05-28)
+ +## Ethical Considerations:
+NVIDIA believes Trustworthy AI is a shared responsibility and we have established policies and practices to enable development for a wide array of AI applications. When downloaded or used in accordance with our terms of service, developers should work with their internal team to ensure this skill meets requirements for the relevant industry and use case and addresses unforeseen product misuse.
+ +(For Release on NVIDIA Platforms Only)
+Please report quality, risk, security vulnerabilities or NVIDIA AI Concerns [here](https://app.intigriti.com/programs/nvidia/nvidiavdp/detail).
diff --git a/skills/nemo-mbridge-perf-moe-comm-overlap/skill.oms.sig b/skills/nemo-mbridge-perf-moe-comm-overlap/skill.oms.sig new file mode 100644 index 0000000000..62a845ebe4 --- /dev/null +++ b/skills/nemo-mbridge-perf-moe-comm-overlap/skill.oms.sig @@ -0,0 +1 @@ +{"mediaType":"application/vnd.dev.sigstore.bundle.v0.3+json","verificationMaterial":{"x509CertificateChain":{"certificates":[{"rawBytes":"MIICgzCCAgmgAwIBAgIUKIyS7SxNteQIiWzK1dWj85E6520wCgYIKoZIzj0EAwMwVTELMAkGA1UEBhMCVVMxGzAZBgNVBAoMEk5WSURJQSBDb3Jwb3JhdGlvbjEpMCcGA1UEAwwgTlZJRElBIEFnZW50IENhcGFiaWxpdGllcyBJQ0EgMDEwHhcNMjYwNDAxMDAwMDAwWhcNMjgwNDIyMTUzMzA5WjBUMQswCQYDVQQGEwJVUzEbMBkGA1UECgwSTlZJRElBIENvcnBvcmF0aW9uMSgwJgYDVQQDDB9OVklESUEgQWdlbnQgU2tpbGxzIFNpZ25pbmcgMDAxMHYwEAYHKoZIzj0CAQYFK4EEACIDYgAEYoRM9bQl/dGlwSRNi6bTpIJUXH8Nv9GciP6LSflJYYMLCc296kpyuTSsk5ddbAWiDcFX3C/ydX3jwc+qCLYP6uHy9XphyLjOQ27Yb2J6rBLVtRBS1mgGco/Gr7fL6ODco4GaMIGXMB0GA1UdDgQWBBRQ/5ZW3nJ6lmo9SVk7I15o7UGmpTAfBgNVHSMEGDAWgBRPGpILxMBBleJSsBGjrMKsby1CgjAMBgNVHRMBAf8EAjAAMA4GA1UdDwEB/wQEAwIHgDA3BggrBgEFBQcBAQQrMCkwJwYIKwYBBQUHMAGGG2h0dHA6Ly9vY3NwLm5kaXMubnZpZGlhLmNvbTAKBggqhkjOPQQDAwNoADBlAjAUygu/GiOCIXrgGr4SmLgeEVDcEitfFUv7ALbvLVGVyMysB3mxmO/uInZfXzWcJZsCMQDxuoxj4ZmO30jhkPIcCxGFCOvnUsnfU3TfGcouYm4M6iRpbKvtVnHPiy4bi6pcKf0="},{"rawBytes":"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"},{"rawBytes":"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"}]},"tlogEntries":[]},"dsseEnvelope":{"payload":"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","payloadType":"application/vnd.in-toto+json","signatures":[{"sig":"MGYCMQDqmJrncOHrfq4OlbR1NpWOTG/7vvNQPJ9ZVDhloZyoYUxhEBeDr+LssQNAs1D5VGkCMQClX79K8gxpq+cZxsUpwQHm39Gc0iAQSVRwQYvyjBWrqiZv25E8+MmO5n5p5ZfSmGc=","keyid":""}]}} \ No newline at end of file From 91d86750f7c868c5b47535076e0abaa4164fa891 Mon Sep 17 00:00:00 2001 From: Chen Cui Date: Thu, 28 May 2026 16:27:00 -0700 Subject: [PATCH 3/6] [skills] fix overlap skill references Signed-off-by: Chen Cui --- docs/training/activation-recomputation.md | 2 +- docs/training/communication-overlap.md | 10 +++++----- docs/training/moe-optimization.md | 2 +- skills/nemo-mbridge-perf-activation-recompute/SKILL.md | 4 ++-- .../nemo-mbridge-perf-activation-recompute/card.yaml | 2 +- skills/perf-cuda-graphs/SKILL.md | 2 +- skills/perf-cuda-graphs/card.yaml | 2 +- skills/perf-memory-tuning/SKILL.md | 6 +++--- 8 files changed, 15 insertions(+), 15 deletions(-) diff --git a/docs/training/activation-recomputation.md b/docs/training/activation-recomputation.md index 3ef8049ef7..b119a64466 100644 --- a/docs/training/activation-recomputation.md +++ b/docs/training/activation-recomputation.md @@ -193,5 +193,5 @@ up-projection modules than to enable blanket full recomputation immediately. - [docs/training/cuda-graphs.md](cuda-graphs.md) - [docs/training/moe-optimization.md](moe-optimization.md) -- [skills/perf-activation-recompute/SKILL.md](../skills/perf-activation-recompute/SKILL.md) — per-module cost/savings data, measured results +- [skills/nemo-mbridge-perf-activation-recompute/SKILL.md](../skills/nemo-mbridge-perf-activation-recompute/SKILL.md) — per-module cost/savings data, measured results - [skills/perf-memory-tuning/SKILL.md](../skills/perf-memory-tuning/SKILL.md) — expandable segments, parallelism resizing, and other memory reduction strategies diff --git a/docs/training/communication-overlap.md b/docs/training/communication-overlap.md index 1e6a779141..85b45b8fe7 100644 --- a/docs/training/communication-overlap.md +++ b/docs/training/communication-overlap.md @@ -8,7 +8,7 @@ to help, and which boundaries are durable across Megatron Bridge. For exact knobs, code anchors, and verification commands, see: - [skills/perf-tp-dp-comm-overlap/SKILL.md](../skills/perf-tp-dp-comm-overlap/SKILL.md) -- [skills/perf-expert-parallel-overlap/SKILL.md](../skills/perf-expert-parallel-overlap/SKILL.md) +- [skills/nemo-mbridge-perf-expert-parallel-overlap/SKILL.md](../skills/nemo-mbridge-perf-expert-parallel-overlap/SKILL.md) ## What It Is @@ -123,7 +123,7 @@ on the actual bottleneck. For config examples and minimal runnable commands, see: - [skills/perf-tp-dp-comm-overlap/SKILL.md](../skills/perf-tp-dp-comm-overlap/SKILL.md) -- [skills/perf-expert-parallel-overlap/SKILL.md](../skills/perf-expert-parallel-overlap/SKILL.md) +- [skills/nemo-mbridge-perf-expert-parallel-overlap/SKILL.md](../skills/nemo-mbridge-perf-expert-parallel-overlap/SKILL.md) ## Expected Metric Changes @@ -159,6 +159,6 @@ correct question is always "which communication path is exposed in this run?" - [docs/training/cuda-graphs.md](cuda-graphs.md) - [docs/training/hierarchical-context-parallel.md](hierarchical-context-parallel.md) - [skills/perf-tp-dp-comm-overlap/SKILL.md](../skills/perf-tp-dp-comm-overlap/SKILL.md) -- [skills/perf-expert-parallel-overlap/SKILL.md](../skills/perf-expert-parallel-overlap/SKILL.md) -- [skills/perf-moe-comm-overlap/SKILL.md](../skills/perf-moe-comm-overlap/SKILL.md) -- [skills/perf-moe-comm-overlap/card.yaml](../skills/perf-moe-comm-overlap/card.yaml) +- [skills/nemo-mbridge-perf-expert-parallel-overlap/SKILL.md](../skills/nemo-mbridge-perf-expert-parallel-overlap/SKILL.md) +- [skills/nemo-mbridge-perf-moe-comm-overlap/SKILL.md](../skills/nemo-mbridge-perf-moe-comm-overlap/SKILL.md) +- [skills/nemo-mbridge-perf-moe-comm-overlap/card.yaml](../skills/nemo-mbridge-perf-moe-comm-overlap/card.yaml) diff --git a/docs/training/moe-optimization.md b/docs/training/moe-optimization.md index e4ad6d95ca..ff3922f45e 100644 --- a/docs/training/moe-optimization.md +++ b/docs/training/moe-optimization.md @@ -11,7 +11,7 @@ For tuning knobs, hardware-specific configs, and benchmark-oriented guidance, se - [skills/perf-moe-long-context/SKILL.md](../skills/perf-moe-long-context/SKILL.md) - [skills/perf-moe-hardware-configs/SKILL.md](../skills/perf-moe-hardware-configs/SKILL.md) - [skills/perf-moe-vlm-training/SKILL.md](../skills/perf-moe-vlm-training/SKILL.md) -- [skills/perf-moe-comm-overlap/SKILL.md](../skills/perf-moe-comm-overlap/SKILL.md) +- [skills/nemo-mbridge-perf-moe-comm-overlap/SKILL.md](../skills/nemo-mbridge-perf-moe-comm-overlap/SKILL.md) ## The Three Walls diff --git a/skills/nemo-mbridge-perf-activation-recompute/SKILL.md b/skills/nemo-mbridge-perf-activation-recompute/SKILL.md index 3eb1570ba9..25739c733a 100644 --- a/skills/nemo-mbridge-perf-activation-recompute/SKILL.md +++ b/skills/nemo-mbridge-perf-activation-recompute/SKILL.md @@ -29,7 +29,7 @@ how many layers via `recompute_num_layers`. 1. **Set `PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True` first** — most borderline OOMs are caused by memory fragmentation, not capacity. This - fixes it at zero cost. See @skills/mbridge-perf-memory-tuning/SKILL.md. + fixes it at zero cost. See @skills/perf-memory-tuning/SKILL.md. 2. Start with `recompute_granularity=selective`, `recompute_modules=[core_attn]` (often already the default in recipes). 3. Add `layernorm` to recompute modules — nearly free compute-wise but saves @@ -129,7 +129,7 @@ Key takeaways: (hidden=28672) is expensive to recompute - Combining `mlp` + `core_attn` is slightly worse than `mlp` alone - For this workload, the actual OOM fix was `PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True` - (memory fragmentation, not capacity). See @skills/mbridge-perf-memory-tuning/SKILL.md. + (memory fragmentation, not capacity). See @skills/perf-memory-tuning/SKILL.md. ## Code Anchors diff --git a/skills/nemo-mbridge-perf-activation-recompute/card.yaml b/skills/nemo-mbridge-perf-activation-recompute/card.yaml index 2f2a1a5832..96f6d8ea36 100644 --- a/skills/nemo-mbridge-perf-activation-recompute/card.yaml +++ b/skills/nemo-mbridge-perf-activation-recompute/card.yaml @@ -92,7 +92,7 @@ recommended_path: first_try: "recompute_granularity=selective, recompute_modules=[core_attn]" if_still_oom: "add layernorm (cheap) or mlp (expensive but saves ~3 GB)" last_resort: "recompute_granularity=full, recompute_method=uniform" - alternative: "see skills/mbridge-perf-memory-tuning/ for VPP tuning and other memory strategies" + alternative: "see skills/perf-memory-tuning/ for VPP tuning and other memory strategies" expected_metric_change: - metric: peak_memory direction: down diff --git a/skills/perf-cuda-graphs/SKILL.md b/skills/perf-cuda-graphs/SKILL.md index 54b8d504ae..530d5df0d1 100644 --- a/skills/perf-cuda-graphs/SKILL.md +++ b/skills/perf-cuda-graphs/SKILL.md @@ -302,7 +302,7 @@ def _delete_cuda_graphs(cuda_graph_helper): recompute (`recompute_granularity="selective"` + `recompute_modules`), disable CUDA graphs, or switch to `local` + `full_iteration`. Enforced in MCore `transformer_config.py:2001-2005`. See also - @skills/perf-activation-recompute/SKILL.md. + @skills/nemo-mbridge-perf-activation-recompute/SKILL.md. 12. **Benchmark numbers are workload-specific**: graph wins are usually real when host overhead is visible, but the exact gain depends on batch shape, diff --git a/skills/perf-cuda-graphs/card.yaml b/skills/perf-cuda-graphs/card.yaml index a2bbae49db..24874b1c49 100644 --- a/skills/perf-cuda-graphs/card.yaml +++ b/skills/perf-cuda-graphs/card.yaml @@ -241,7 +241,7 @@ failure_modes: - name: full_recompute_with_te_scoped symptom: "AssertionError: full recompute is only supported with full iteration CUDA graph" likely_cause: recompute_granularity=full with any TE-scoped graph (attn, mlp, moe_router, etc.). Common on FP8 CS configs that default to cuda_graph_impl=transformer_engine + scope=mlp. - fix: use recompute_granularity=selective with recompute_modules, or disable CUDA graphs (cuda_graph_impl=none), or switch to cuda_graph_impl=local + cuda_graph_scope=full_iteration. See skills/perf-activation-recompute/SKILL.md. + fix: use recompute_granularity=selective with recompute_modules, or disable CUDA graphs (cuda_graph_impl=none), or switch to cuda_graph_impl=local + cuda_graph_scope=full_iteration. See skills/nemo-mbridge-perf-activation-recompute/SKILL.md. - name: packed_sequences_with_te_scoped symptom: "AssertionError: CUDA graph accepts only Tensor inputs. packed_seq_params excluded" likely_cause: packed_sequence=True passes a non-Tensor packed_seq_params input into TE-scoped capture diff --git a/skills/perf-memory-tuning/SKILL.md b/skills/perf-memory-tuning/SKILL.md index df86807b1b..7bc5c1033f 100644 --- a/skills/perf-memory-tuning/SKILL.md +++ b/skills/perf-memory-tuning/SKILL.md @@ -53,7 +53,7 @@ When a training run OOMs or is close to the memory limit: fragmentation-induced OOM with zero performance cost. Most Slurm launch templates already include it. 2. **Add selective activation recompute** (`recompute_modules=[core_attn]`) if - not already enabled. See @skills/perf-activation-recompute/SKILL.md. + not already enabled. See @skills/nemo-mbridge-perf-activation-recompute/SKILL.md. 3. **Avoid increasing TP** as a memory fix — doubling TP dramatically increases NVLink all-reduce volume and often kills throughput (-28% on Llama3 70B). 4. **Avoid increasing PP at the cost of DP** — halving DP doubles gradient @@ -95,7 +95,7 @@ If the model genuinely does not fit (not fragmentation), adjust parallelism: ### Activation recompute -See @skills/perf-activation-recompute/SKILL.md for full details. +See @skills/nemo-mbridge-perf-activation-recompute/SKILL.md for full details. ### CPU offloading @@ -174,7 +174,7 @@ offloading.` This approach is blocked for any model using PP > 1. Selective activation recompute with `mlp` saved ~3 GB peak memory but cost ~16% GPU utilization on this workload. See -@skills/perf-activation-recompute/SKILL.md for full results. +@skills/nemo-mbridge-perf-activation-recompute/SKILL.md for full results. ## Code Anchors From c24083070d91eba76f4ab5803631d6d4e1d93ed5 Mon Sep 17 00:00:00 2001 From: Chen Cui Date: Thu, 28 May 2026 17:08:04 -0700 Subject: [PATCH 4/6] [skills] keep signing batch scoped Signed-off-by: Chen Cui --- skills/perf-cuda-graphs/SKILL.md | 2 +- skills/perf-cuda-graphs/card.yaml | 2 +- skills/perf-memory-tuning/SKILL.md | 6 +++--- 3 files changed, 5 insertions(+), 5 deletions(-) diff --git a/skills/perf-cuda-graphs/SKILL.md b/skills/perf-cuda-graphs/SKILL.md index 530d5df0d1..54b8d504ae 100644 --- a/skills/perf-cuda-graphs/SKILL.md +++ b/skills/perf-cuda-graphs/SKILL.md @@ -302,7 +302,7 @@ def _delete_cuda_graphs(cuda_graph_helper): recompute (`recompute_granularity="selective"` + `recompute_modules`), disable CUDA graphs, or switch to `local` + `full_iteration`. Enforced in MCore `transformer_config.py:2001-2005`. See also - @skills/nemo-mbridge-perf-activation-recompute/SKILL.md. + @skills/perf-activation-recompute/SKILL.md. 12. **Benchmark numbers are workload-specific**: graph wins are usually real when host overhead is visible, but the exact gain depends on batch shape, diff --git a/skills/perf-cuda-graphs/card.yaml b/skills/perf-cuda-graphs/card.yaml index 24874b1c49..a2bbae49db 100644 --- a/skills/perf-cuda-graphs/card.yaml +++ b/skills/perf-cuda-graphs/card.yaml @@ -241,7 +241,7 @@ failure_modes: - name: full_recompute_with_te_scoped symptom: "AssertionError: full recompute is only supported with full iteration CUDA graph" likely_cause: recompute_granularity=full with any TE-scoped graph (attn, mlp, moe_router, etc.). Common on FP8 CS configs that default to cuda_graph_impl=transformer_engine + scope=mlp. - fix: use recompute_granularity=selective with recompute_modules, or disable CUDA graphs (cuda_graph_impl=none), or switch to cuda_graph_impl=local + cuda_graph_scope=full_iteration. See skills/nemo-mbridge-perf-activation-recompute/SKILL.md. + fix: use recompute_granularity=selective with recompute_modules, or disable CUDA graphs (cuda_graph_impl=none), or switch to cuda_graph_impl=local + cuda_graph_scope=full_iteration. See skills/perf-activation-recompute/SKILL.md. - name: packed_sequences_with_te_scoped symptom: "AssertionError: CUDA graph accepts only Tensor inputs. packed_seq_params excluded" likely_cause: packed_sequence=True passes a non-Tensor packed_seq_params input into TE-scoped capture diff --git a/skills/perf-memory-tuning/SKILL.md b/skills/perf-memory-tuning/SKILL.md index 7bc5c1033f..df86807b1b 100644 --- a/skills/perf-memory-tuning/SKILL.md +++ b/skills/perf-memory-tuning/SKILL.md @@ -53,7 +53,7 @@ When a training run OOMs or is close to the memory limit: fragmentation-induced OOM with zero performance cost. Most Slurm launch templates already include it. 2. **Add selective activation recompute** (`recompute_modules=[core_attn]`) if - not already enabled. See @skills/nemo-mbridge-perf-activation-recompute/SKILL.md. + not already enabled. See @skills/perf-activation-recompute/SKILL.md. 3. **Avoid increasing TP** as a memory fix — doubling TP dramatically increases NVLink all-reduce volume and often kills throughput (-28% on Llama3 70B). 4. **Avoid increasing PP at the cost of DP** — halving DP doubles gradient @@ -95,7 +95,7 @@ If the model genuinely does not fit (not fragmentation), adjust parallelism: ### Activation recompute -See @skills/nemo-mbridge-perf-activation-recompute/SKILL.md for full details. +See @skills/perf-activation-recompute/SKILL.md for full details. ### CPU offloading @@ -174,7 +174,7 @@ offloading.` This approach is blocked for any model using PP > 1. Selective activation recompute with `mlp` saved ~3 GB peak memory but cost ~16% GPU utilization on this workload. See -@skills/nemo-mbridge-perf-activation-recompute/SKILL.md for full results. +@skills/perf-activation-recompute/SKILL.md for full results. ## Code Anchors From 6fe590a5009d6d7277c2be848db05c0a5bb8acb4 Mon Sep 17 00:00:00 2001 From: Chen Cui Date: Thu, 28 May 2026 17:25:04 -0700 Subject: [PATCH 5/6] [skills] keep activation recompute references final Signed-off-by: Chen Cui --- docs/training/activation-recomputation.md | 2 +- skills/nemo-mbridge-perf-activation-recompute/SKILL.md | 4 ++-- skills/nemo-mbridge-perf-activation-recompute/card.yaml | 2 +- 3 files changed, 4 insertions(+), 4 deletions(-) diff --git a/docs/training/activation-recomputation.md b/docs/training/activation-recomputation.md index b119a64466..0e1d47b5ed 100644 --- a/docs/training/activation-recomputation.md +++ b/docs/training/activation-recomputation.md @@ -194,4 +194,4 @@ up-projection modules than to enable blanket full recomputation immediately. - [docs/training/cuda-graphs.md](cuda-graphs.md) - [docs/training/moe-optimization.md](moe-optimization.md) - [skills/nemo-mbridge-perf-activation-recompute/SKILL.md](../skills/nemo-mbridge-perf-activation-recompute/SKILL.md) — per-module cost/savings data, measured results -- [skills/perf-memory-tuning/SKILL.md](../skills/perf-memory-tuning/SKILL.md) — expandable segments, parallelism resizing, and other memory reduction strategies +- [skills/nemo-mbridge-perf-memory-tuning/SKILL.md](../skills/nemo-mbridge-perf-memory-tuning/SKILL.md) — expandable segments, parallelism resizing, and other memory reduction strategies diff --git a/skills/nemo-mbridge-perf-activation-recompute/SKILL.md b/skills/nemo-mbridge-perf-activation-recompute/SKILL.md index 25739c733a..c375a86704 100644 --- a/skills/nemo-mbridge-perf-activation-recompute/SKILL.md +++ b/skills/nemo-mbridge-perf-activation-recompute/SKILL.md @@ -29,7 +29,7 @@ how many layers via `recompute_num_layers`. 1. **Set `PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True` first** — most borderline OOMs are caused by memory fragmentation, not capacity. This - fixes it at zero cost. See @skills/perf-memory-tuning/SKILL.md. + fixes it at zero cost. See @skills/nemo-mbridge-perf-memory-tuning/SKILL.md. 2. Start with `recompute_granularity=selective`, `recompute_modules=[core_attn]` (often already the default in recipes). 3. Add `layernorm` to recompute modules — nearly free compute-wise but saves @@ -129,7 +129,7 @@ Key takeaways: (hidden=28672) is expensive to recompute - Combining `mlp` + `core_attn` is slightly worse than `mlp` alone - For this workload, the actual OOM fix was `PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True` - (memory fragmentation, not capacity). See @skills/perf-memory-tuning/SKILL.md. + (memory fragmentation, not capacity). See @skills/nemo-mbridge-perf-memory-tuning/SKILL.md. ## Code Anchors diff --git a/skills/nemo-mbridge-perf-activation-recompute/card.yaml b/skills/nemo-mbridge-perf-activation-recompute/card.yaml index 96f6d8ea36..f5df90608c 100644 --- a/skills/nemo-mbridge-perf-activation-recompute/card.yaml +++ b/skills/nemo-mbridge-perf-activation-recompute/card.yaml @@ -92,7 +92,7 @@ recommended_path: first_try: "recompute_granularity=selective, recompute_modules=[core_attn]" if_still_oom: "add layernorm (cheap) or mlp (expensive but saves ~3 GB)" last_resort: "recompute_granularity=full, recompute_method=uniform" - alternative: "see skills/perf-memory-tuning/ for VPP tuning and other memory strategies" + alternative: "see skills/nemo-mbridge-perf-memory-tuning/ for VPP tuning and other memory strategies" expected_metric_change: - metric: peak_memory direction: down From 1921fa5871bda31a3904ed3744c22002a9146101 Mon Sep 17 00:00:00 2001 From: nvskills-svc-account Date: Fri, 29 May 2026 00:36:37 +0000 Subject: [PATCH 6/6] Attach NVSkills validation signatures Signed-off-by: nvskills-svc-account --- .../BENCHMARK.md | 4 ++-- .../skill-card.md | 9 +++++---- .../skill.oms.sig | 2 +- .../BENCHMARK.md | 4 ++-- .../skill-card.md | 9 +++++---- .../skill.oms.sig | 2 +- .../BENCHMARK.md | 8 ++++---- .../skill-card.md | 16 +++++++--------- .../skill.oms.sig | 2 +- 9 files changed, 28 insertions(+), 28 deletions(-) diff --git a/skills/nemo-mbridge-perf-activation-recompute/BENCHMARK.md b/skills/nemo-mbridge-perf-activation-recompute/BENCHMARK.md index e2ae5dd1f4..979410a496 100644 --- a/skills/nemo-mbridge-perf-activation-recompute/BENCHMARK.md +++ b/skills/nemo-mbridge-perf-activation-recompute/BENCHMARK.md @@ -7,7 +7,7 @@ This benchmark summarizes 3-Tier Evaluation from NVSkills-Eval results for the s ## Evaluation Summary - Skill: `nemo-mbridge-perf-activation-recompute` -- Evaluation date: 2026-05-28 +- Evaluation date: 2026-05-29 - NVSkills-Eval profile: `external` - Overall verdict: FAIL - Tier 3 live agent evaluation: not available in this report @@ -40,7 +40,7 @@ Tier 3 dimension rollup was not available in this report. ## Tier 1: Static Validation Summary -Tier 1 validation passed with observations. NVSkills-Eval ran 9 checks and found 10 total findings. +Tier 1 validation reported findings. NVSkills-Eval ran 9 checks and found 13 total findings. Top findings: diff --git a/skills/nemo-mbridge-perf-activation-recompute/skill-card.md b/skills/nemo-mbridge-perf-activation-recompute/skill-card.md index cfd22d1c7a..cefe25637f 100644 --- a/skills/nemo-mbridge-perf-activation-recompute/skill-card.md +++ b/skills/nemo-mbridge-perf-activation-recompute/skill-card.md @@ -9,7 +9,7 @@ NVIDIA
### License/Terms of Use:
Apache 2.0
## Use Case:
-Developers and engineers configuring activation recompute strategies in Megatron Bridge to reduce GPU memory consumption during large language model training, or diagnosing OOM and performance regressions caused by recompute settings.
+Developers and engineers reducing GPU memory pressure during LLM training by configuring activation recompute strategies in Megatron Bridge, or diagnosing OOM failures and performance regressions related to recompute settings.
### Deployment Geography for Use:
Global
@@ -21,16 +21,17 @@ Mitigation: Review and scan skill before deployment.
## Reference(s):
- [Activation Recomputation Documentation](docs/training/activation-recomputation.md)
- [Performance Tuning Guide](docs/performance-guide.md)
+- [Megatron Bridge Documentation](https://docs.nvidia.com/nemo/megatron-bridge/latest/)
## Skill Output:
**Output Type(s):** [Configuration instructions, Shell commands, Analysis]
-**Output Format:** [Markdown with inline Python and bash code blocks]
+**Output Format:** [Markdown with inline code blocks]
**Output Parameters:** [1D]
**Other Properties Related to Output:** [None]
## Evaluation Tasks:
-NVSkills-Eval 3-Tier evaluation (external profile) assessing skill safety, discoverability, effectiveness, and usefulness for agents before publication.
+3-Tier NVSkills-Eval evaluation covering static validation, deduplication, and live agent testing for safety, correctness, discoverability, effectiveness, and efficiency.
## Evaluation Metrics Used:
Reported benchmark dimensions:
@@ -43,7 +44,7 @@ Reported benchmark dimensions:
## Skill Version(s):
-v0.2.0rc6-1465-gcdb9c50d (source: git describe)
+v0.2.0rc6-1469-g6fe590a5 (source: git tag)
## Ethical Considerations:
NVIDIA believes Trustworthy AI is a shared responsibility and we have established policies and practices to enable development for a wide array of AI applications. When downloaded or used in accordance with our terms of service, developers should work with their internal team to ensure this skill meets requirements for the relevant industry and use case and addresses unforeseen product misuse.
diff --git a/skills/nemo-mbridge-perf-activation-recompute/skill.oms.sig b/skills/nemo-mbridge-perf-activation-recompute/skill.oms.sig index 72804f953b..ad7d628473 100644 --- a/skills/nemo-mbridge-perf-activation-recompute/skill.oms.sig +++ b/skills/nemo-mbridge-perf-activation-recompute/skill.oms.sig @@ -1 +1 @@ -{"mediaType":"application/vnd.dev.sigstore.bundle.v0.3+json","verificationMaterial":{"x509CertificateChain":{"certificates":[{"rawBytes":"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"},{"rawBytes":"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"},{"rawBytes":"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"}]},"tlogEntries":[]},"dsseEnvelope":{"payload":"ewogICJfdHlwZSI6ICJodHRwczovL2luLXRvdG8uaW8vU3RhdGVtZW50L3YxIiwKICAic3ViamVjdCI6IFsKICAgIHsKICAgICAgIm5hbWUiOiAibmVtby1tYnJpZGdlLXBlcmYtYWN0aXZhdGlvbi1yZWNvbXB1dGUiLAogICAgICAiZGlnZXN0IjogewogICAgICAgICJzaGEyNTYiOiAiNWJkZWZkN2I5YzdkNGZjYjM4NDNmOTExODE4MDA1ZTBjOTE5MDliMjFjNGY2MDYwNjRlYmUxZTBmZDM1NzRlNSIKICAgICAgfQogICAgfQogIF0sCiAgInByZWRpY2F0ZVR5cGUiOiAiaHR0cHM6Ly9tb2RlbF9zaWduaW5nL3NpZ25hdHVyZS92MS4wIiwKICAicHJlZGljYXRlIjogewogICAgInJlc291cmNlcyI6IFsKICAgICAgewogICAgICAgICJhbGdvcml0aG0iOiAic2hhMjU2IiwKICAgICAgICAibmFtZSI6ICJCRU5DSE1BUksubWQiLAogICAgICAgICJkaWdlc3QiOiAiMTE4YTk5MzllNjUwNGFlYzU4YTgxOTU4NmJhOTYwZGQ2YTZkZDEyNWY2Njg3OTMxODY3MjU1YzVmYjY2MmZjMSIKICAgICAgfSwKICAgICAgewogICAgICAgICJhbGdvcml0aG0iOiAic2hhMjU2IiwKICAgICAgICAibmFtZSI6ICJTS0lMTC5tZCIsCiAgICAgICAgImRpZ2VzdCI6ICJkMjhmZGViMzEwMzljZDJlZmQ3ZWQxM2Q5NDA3ZDgzYWIyM2E4OWMyOWMyMzY3YzdkZGQwZTJkMDA1MjExNWE3IgogICAgICB9LAogICAgICB7CiAgICAgICAgImFsZ29yaXRobSI6ICJzaGEyNTYiLAogICAgICAgICJuYW1lIjogImNhcmQueWFtbCIsCiAgICAgICAgImRpZ2VzdCI6ICIyNTNjMGZhOWJjMWQwOThjMWM4NjJjZGU0YWNiOThlYjRjYTM5OWNhM2RiZWY3NmFkMTFkYTc0MWEzODNlNjcxIgogICAgICB9LAogICAgICB7CiAgICAgICAgImFsZ29yaXRobSI6ICJzaGEyNTYiLAogICAgICAgICJuYW1lIjogImV2YWxzL2V2YWxzLmpzb24iLAogICAgICAgICJkaWdlc3QiOiAiMzc1MTdlNWYzZGM2NjgxOWY2MWY1YTdiYjhhY2UxOTIxMjgyNDE1ZjEwNTUxZDJkZWZhNWMzZWIwOTg1YjU3MCIKICAgICAgfSwKICAgICAgewogICAgICAgICJhbGdvcml0aG0iOiAic2hhMjU2IiwKICAgICAgICAibmFtZSI6ICJza2lsbC1jYXJkLm1kIiwKICAgICAgICAiZGlnZXN0IjogIjY5YzNlOTA3M2M4YTVlYjI4MmEyZWM5NDYyM2FjNGEzMmM4MTk0YTg5YTViMDRlN2MyM2FhMDljYjYyYTg1OTIiCiAgICAgIH0KICAgIF0sCiAgICAic2VyaWFsaXphdGlvbiI6IHsKICAgICAgImhhc2hfdHlwZSI6ICJzaGEyNTYiLAogICAgICAiaWdub3JlX3BhdGhzIjogWwogICAgICAgICIuZ2l0aWdub3JlIiwKICAgICAgICAiLmdpdCIsCiAgICAgICAgIi5naXRhdHRyaWJ1dGVzIiwKICAgICAgICAiLmdpdGh1YiIKICAgICAgXSwKICAgICAgImFsbG93X3N5bWxpbmtzIjogZmFsc2UsCiAgICAgICJtZXRob2QiOiAiZmlsZXMiCiAgICB9CiAgfQp9","payloadType":"application/vnd.in-toto+json","signatures":[{"sig":"MGUCMQD6eiJvxXq6hfdLhIrDVw+3soHxglRo8sb2LIjjuFp7LFrDtBaH04LNwR+2y40Jfm4CMEW6ZQZ3//bMyVqzB1V8q1EW+MOwEhjlsZh4S6QIZpJCP42j//lLgsegHxwBAhmgXw==","keyid":""}]}} \ No newline at end of file +{"mediaType":"application/vnd.dev.sigstore.bundle.v0.3+json","verificationMaterial":{"x509CertificateChain":{"certificates":[{"rawBytes":"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"},{"rawBytes":"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"},{"rawBytes":"MIICRTCCAcygAwIBAgIUeJdY3rV86EdvFmG7L8LJBsyQFYkwCgYIKoZIzj0EAwMwUTELMAkGA1UEBhMCVVMxGzAZBgNVBAoMEk5WSURJQSBDb3Jwb3JhdGlvbjElMCMGA1UEAwwcTlZJRElBIEFnZW50IENhcGFiaWxpdGllcyBDQTAgFw0yNjA0MDEwMDAwMDBaGA85OTk5MTIzMTIzNTk1OVowUTELMAkGA1UEBhMCVVMxGzAZBgNVBAoMEk5WSURJQSBDb3Jwb3JhdGlvbjElMCMGA1UEAwwcTlZJRElBIEFnZW50IENhcGFiaWxpdGllcyBDQTB2MBAGByqGSM49AgEGBSuBBAAiA2IABAYpiXCDjJ9NT2eSDhyHJVSw1Tbze18cGG2F/578oWvHxg23eQAhNRYdq88i1iOshZSO6C29doKui5Xpmo/7Ctw9Sx4PP2RzOmIuOLCuTdNtKcTRwi4GEsd5BAFvWj42M6NjMGEwHQYDVR0OBBYEFItnoAjjfuCEUvzyvWyI2vOGvwPjMB8GA1UdIwQYMBaAFItnoAjjfuCEUvzyvWyI2vOGvwPjMA8GA1UdEwEB/wQFMAMBAf8wDgYDVR0PAQH/BAQDAgEGMAoGCCqGSM49BAMDA2cAMGQCMCwtAjWLaNwgGWNCgdyNoTyvNhqWRECRJV2r3+7w8g0PL6NHLOsbkgE09BH95h8XlgIwTaQmbbUh2ChAJ5TA1wRiVDnCcvbzHlZl2jM2FcwQQZlk19LOAbyGMRixbu2Ww/rj"}]},"tlogEntries":[]},"dsseEnvelope":{"payload":"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","payloadType":"application/vnd.in-toto+json","signatures":[{"sig":"MGUCMAp5cStPxWc7M319y40/FCLEE/6OcYEjnBjIgI423c6e1dtNYjuFNwwvA87ebc4wAgIxAPFyHCNvOn9vLL2/slR8L6pEXG+ZccygYuyN62W9BW+wmWtyb+nszsqdCcTuKAFenw==","keyid":""}]}} \ No newline at end of file diff --git a/skills/nemo-mbridge-perf-expert-parallel-overlap/BENCHMARK.md b/skills/nemo-mbridge-perf-expert-parallel-overlap/BENCHMARK.md index e0f009499d..cdd4142400 100644 --- a/skills/nemo-mbridge-perf-expert-parallel-overlap/BENCHMARK.md +++ b/skills/nemo-mbridge-perf-expert-parallel-overlap/BENCHMARK.md @@ -7,7 +7,7 @@ This benchmark summarizes 3-Tier Evaluation from NVSkills-Eval results for the s ## Evaluation Summary - Skill: `nemo-mbridge-perf-expert-parallel-overlap` -- Evaluation date: 2026-05-28 +- Evaluation date: 2026-05-29 - NVSkills-Eval profile: `external` - Overall verdict: FAIL - Tier 3 live agent evaluation: not available in this report @@ -40,7 +40,7 @@ Tier 3 dimension rollup was not available in this report. ## Tier 1: Static Validation Summary -Tier 1 validation passed with observations. NVSkills-Eval ran 9 checks and found 11 total findings. +Tier 1 validation reported findings. NVSkills-Eval ran 9 checks and found 14 total findings. Top findings: diff --git a/skills/nemo-mbridge-perf-expert-parallel-overlap/skill-card.md b/skills/nemo-mbridge-perf-expert-parallel-overlap/skill-card.md index b0096b6df1..2510d98844 100644 --- a/skills/nemo-mbridge-perf-expert-parallel-overlap/skill-card.md +++ b/skills/nemo-mbridge-perf-expert-parallel-overlap/skill-card.md @@ -9,7 +9,7 @@ NVIDIA
### License/Terms of Use:
Apache 2.0
## Use Case:
-Developers and engineers enabling expert-parallel communication overlap to hide dispatch/combine latency in MoE model training, or diagnosing throughput regressions related to EP overlap configuration changes in Megatron-Bridge.
+Developers and engineers enabling MoE expert-parallel communication overlap to optimize distributed training throughput for Mixture-of-Experts models in Megatron-Bridge.
### Deployment Geography for Use:
Global
@@ -19,12 +19,13 @@ Risk: Review before execution as proposals could introduce incorrect or misleadi Mitigation: Review and scan skill before deployment.
## Reference(s):
+- [Communication Overlap Documentation](docs/training/communication-overlap.md)
- [Performance Tuning Guide](docs/performance-guide.md)
-- [Megatron-Bridge GitHub Repository](https://github.com/NVIDIA-NeMo/Megatron-Bridge)
+- [Megatron-Bridge Documentation](https://docs.nvidia.com/nemo/megatron-bridge/latest/)
## Skill Output:
-**Output Type(s):** [Shell commands, Configuration instructions, Analysis]
+**Output Type(s):** [Configuration instructions, Shell commands, Analysis]
**Output Format:** [Markdown with inline Python and bash code blocks]
**Output Parameters:** [1D]
**Other Properties Related to Output:** [None]
@@ -40,7 +41,7 @@ Reported benchmark dimensions:
## Skill Version(s):
-v0.2.0rc6-1465-gcdb9c50d (source: git describe)
+v0.2.0rc6-1469-g6fe590a5 (source: git describe, committed 2026-05-28)
## Ethical Considerations:
NVIDIA believes Trustworthy AI is a shared responsibility and we have established policies and practices to enable development for a wide array of AI applications. When downloaded or used in accordance with our terms of service, developers should work with their internal team to ensure this skill meets requirements for the relevant industry and use case and addresses unforeseen product misuse.
diff --git a/skills/nemo-mbridge-perf-expert-parallel-overlap/skill.oms.sig b/skills/nemo-mbridge-perf-expert-parallel-overlap/skill.oms.sig index afaa5c29af..53beeda3a2 100644 --- a/skills/nemo-mbridge-perf-expert-parallel-overlap/skill.oms.sig +++ b/skills/nemo-mbridge-perf-expert-parallel-overlap/skill.oms.sig @@ -1 +1 @@ -{"mediaType":"application/vnd.dev.sigstore.bundle.v0.3+json","verificationMaterial":{"x509CertificateChain":{"certificates":[{"rawBytes":"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"},{"rawBytes":"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"},{"rawBytes":"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"}]},"tlogEntries":[]},"dsseEnvelope":{"payload":"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","payloadType":"application/vnd.in-toto+json","signatures":[{"sig":"MGQCMEg6td5XT4+jcUKtqo0MLG/CgEVPEuMmkOVXYGvbF/37U+p76zzEKU5r9arNe1ov7gIwQW6InESVYD/M/kNfBo74eZtA/HG9F5bHGAPR88Jw+yPujbAYVjeBrebtgLANRYb7","keyid":""}]}} \ No newline at end of file +{"mediaType":"application/vnd.dev.sigstore.bundle.v0.3+json","verificationMaterial":{"x509CertificateChain":{"certificates":[{"rawBytes":"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"},{"rawBytes":"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"},{"rawBytes":"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"}]},"tlogEntries":[]},"dsseEnvelope":{"payload":"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","payloadType":"application/vnd.in-toto+json","signatures":[{"sig":"MGUCMBvUmyXebAGy0UC1zrIw6R3zUkIXJV5N/RY9MiuOvOXlhWBtortXrO+fqqp0+9zutQIxAIKarBfEZRRiq9uL0yVAHDAoq6hNsaIHuYvJcP25Nk/aXTSgzsUrLTWH+Vjcm4bxAg==","keyid":""}]}} \ No newline at end of file diff --git a/skills/nemo-mbridge-perf-moe-comm-overlap/BENCHMARK.md b/skills/nemo-mbridge-perf-moe-comm-overlap/BENCHMARK.md index c08ed4b807..d242bab13d 100644 --- a/skills/nemo-mbridge-perf-moe-comm-overlap/BENCHMARK.md +++ b/skills/nemo-mbridge-perf-moe-comm-overlap/BENCHMARK.md @@ -7,9 +7,9 @@ This benchmark summarizes 3-Tier Evaluation from NVSkills-Eval results for the s ## Evaluation Summary - Skill: `nemo-mbridge-perf-moe-comm-overlap` -- Evaluation date: 2026-05-28 +- Evaluation date: 2026-05-29 - NVSkills-Eval profile: `external` -- Overall verdict: PASS +- Overall verdict: FAIL - Tier 3 live agent evaluation: not available in this report ## Agents Used @@ -40,7 +40,7 @@ Tier 3 dimension rollup was not available in this report. ## Tier 1: Static Validation Summary -Tier 1 validation passed with observations. NVSkills-Eval ran 9 checks and found 10 total findings. +Tier 1 validation reported findings. NVSkills-Eval ran 9 checks and found 14 total findings. Top findings: @@ -61,4 +61,4 @@ Notable observations: ## Publication Recommendation -The skill is suitable to proceed toward NVSkills-Eval publication based on this benchmark. Skill owners should keep this file with the skill and refresh it when the evaluation dataset, skill behavior, or target agents materially change. +The skill should be reviewed before NVSkills-Eval publication. Skill owners should address the findings above and rerun NVSkills-Eval to refresh this benchmark. diff --git a/skills/nemo-mbridge-perf-moe-comm-overlap/skill-card.md b/skills/nemo-mbridge-perf-moe-comm-overlap/skill-card.md index 53a23f0e6b..7b7151aee7 100644 --- a/skills/nemo-mbridge-perf-moe-comm-overlap/skill-card.md +++ b/skills/nemo-mbridge-perf-moe-comm-overlap/skill-card.md @@ -1,5 +1,5 @@ ## Description:
-MoE expert-parallel communication overlap in Megatron Bridge. Covers dispatch/combine overlap, flex dispatcher backends, and expert wgrad scheduling.
+MoE expert-parallel communication overlap in Megatron Bridge, covering dispatch/combine overlap, flex dispatcher backends, and expert wgrad scheduling.
This skill is ready for commercial/non-commercial use.
@@ -9,7 +9,7 @@ NVIDIA
### License/Terms of Use:
Apache 2.0
## Use Case:
-Developers and engineers tuning MoE communication overlap for expert-parallel training in Megatron Bridge, or tracing throughput regressions to comm-overlap configuration changes.
+Developers and engineers tuning MoE expert-parallel communication overlap to improve training throughput in Megatron Bridge workloads with expert parallelism.
### Deployment Geography for Use:
Global
@@ -19,20 +19,18 @@ Risk: Review before execution as proposals could introduce incorrect or misleadi Mitigation: Review and scan skill before deployment.
## Reference(s):
-- [Communication Overlap Documentation](docs/training/communication-overlap.md)
+- [Communication Overlap Guide](docs/training/communication-overlap.md)
- [Performance Tuning Guide](docs/performance-guide.md)
-- [Parallelisms Documentation](docs/parallelisms.md)
+- [Comm Overlap Validation Source](src/megatron/bridge/training/comm_overlap.py)
+- [Flex Dispatcher Backend Source](src/megatron/bridge/training/flex_dispatcher_backend.py)
## Skill Output:
-**Output Type(s):** [Configuration instructions, Shell commands, Analysis]
+**Output Type(s):** [Configuration instructions, Shell commands]
**Output Format:** [Markdown with inline code blocks]
**Output Parameters:** [1D]
**Other Properties Related to Output:** [None]
-## Evaluation Tasks:
-NVSkills-Eval 3-Tier evaluation with external profile. Tier 1 static validation (9 checks), Tier 2 deduplication (2 checks). Tier 3 live agent evaluation not available.
- ## Evaluation Metrics Used:
Reported benchmark dimensions:
- Security: Checks whether skill-assisted execution avoids unsafe behavior such as secret leakage, destructive commands, or unauthorized access.
@@ -44,7 +42,7 @@ Reported benchmark dimensions:
## Skill Version(s):
-cdb9c50d (source: git SHA, committed 2026-05-28)
+v0.2.0rc6-1469-g6fe590a5 (source: git describe)
## Ethical Considerations:
NVIDIA believes Trustworthy AI is a shared responsibility and we have established policies and practices to enable development for a wide array of AI applications. When downloaded or used in accordance with our terms of service, developers should work with their internal team to ensure this skill meets requirements for the relevant industry and use case and addresses unforeseen product misuse.
diff --git a/skills/nemo-mbridge-perf-moe-comm-overlap/skill.oms.sig b/skills/nemo-mbridge-perf-moe-comm-overlap/skill.oms.sig index 62a845ebe4..543f6025f2 100644 --- a/skills/nemo-mbridge-perf-moe-comm-overlap/skill.oms.sig +++ b/skills/nemo-mbridge-perf-moe-comm-overlap/skill.oms.sig @@ -1 +1 @@ -{"mediaType":"application/vnd.dev.sigstore.bundle.v0.3+json","verificationMaterial":{"x509CertificateChain":{"certificates":[{"rawBytes":"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"},{"rawBytes":"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"},{"rawBytes":"MIICRTCCAcygAwIBAgIUeJdY3rV86EdvFmG7L8LJBsyQFYkwCgYIKoZIzj0EAwMwUTELMAkGA1UEBhMCVVMxGzAZBgNVBAoMEk5WSURJQSBDb3Jwb3JhdGlvbjElMCMGA1UEAwwcTlZJRElBIEFnZW50IENhcGFiaWxpdGllcyBDQTAgFw0yNjA0MDEwMDAwMDBaGA85OTk5MTIzMTIzNTk1OVowUTELMAkGA1UEBhMCVVMxGzAZBgNVBAoMEk5WSURJQSBDb3Jwb3JhdGlvbjElMCMGA1UEAwwcTlZJRElBIEFnZW50IENhcGFiaWxpdGllcyBDQTB2MBAGByqGSM49AgEGBSuBBAAiA2IABAYpiXCDjJ9NT2eSDhyHJVSw1Tbze18cGG2F/578oWvHxg23eQAhNRYdq88i1iOshZSO6C29doKui5Xpmo/7Ctw9Sx4PP2RzOmIuOLCuTdNtKcTRwi4GEsd5BAFvWj42M6NjMGEwHQYDVR0OBBYEFItnoAjjfuCEUvzyvWyI2vOGvwPjMB8GA1UdIwQYMBaAFItnoAjjfuCEUvzyvWyI2vOGvwPjMA8GA1UdEwEB/wQFMAMBAf8wDgYDVR0PAQH/BAQDAgEGMAoGCCqGSM49BAMDA2cAMGQCMCwtAjWLaNwgGWNCgdyNoTyvNhqWRECRJV2r3+7w8g0PL6NHLOsbkgE09BH95h8XlgIwTaQmbbUh2ChAJ5TA1wRiVDnCcvbzHlZl2jM2FcwQQZlk19LOAbyGMRixbu2Ww/rj"}]},"tlogEntries":[]},"dsseEnvelope":{"payload":"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","payloadType":"application/vnd.in-toto+json","signatures":[{"sig":"MGYCMQDqmJrncOHrfq4OlbR1NpWOTG/7vvNQPJ9ZVDhloZyoYUxhEBeDr+LssQNAs1D5VGkCMQClX79K8gxpq+cZxsUpwQHm39Gc0iAQSVRwQYvyjBWrqiZv25E8+MmO5n5p5ZfSmGc=","keyid":""}]}} \ No newline at end of file +{"mediaType":"application/vnd.dev.sigstore.bundle.v0.3+json","verificationMaterial":{"x509CertificateChain":{"certificates":[{"rawBytes":"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"},{"rawBytes":"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"},{"rawBytes":"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"}]},"tlogEntries":[]},"dsseEnvelope":{"payload":"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","payloadType":"application/vnd.in-toto+json","signatures":[{"sig":"MGYCMQC8qFsDDWi1zvBqAn6A3BN9wtb2lvxESGA/nlI79xykM8NMIAk4oI4Bo4Y+tAyvv5cCMQCxxb1FWqIpz4LYMLkzH2AxLHMn6fX33OV0qfJdpt9hSvT7E+zhZHNAGoJ461YHgCI=","keyid":""}]}} \ No newline at end of file