メニュー
Validate GRPO training configs against 36 cross-framework bug avoidance rules (19 MUST DO + 17 MUST NOT). Covers 7 frameworks (DeepSpeed, Megatron, vLLM, verl, SGLang, rLLM, MindIE), 7 pattern classes, and MoE NaN avoidance.
Codex または Claude でインストール この Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。
Debug and diagnose GRPO training issues across verl, Megatron, DeepSpeed, and RTX 4090. Symptom-to-bug mapping, NaN debugging, Muon clipping avoidance, and CUDA stream safety analysis.
Navigate SGLang source code. Quick reference for subsystem locations, key classes, and data flow.
Navigate mechanistic interpretability concepts, tools, and SAE architectures. Quick reference for SAE variants, feature steering methods, circuit discovery, and trust calibration.
Navigate model merging strategies and theory. Quick reference for Task Arithmetic, DARE, TIES merging methods, alpha selection, and RTX 4090 validation results.
Navigate RDMA, InfiniBand, RoCE v2, AI cluster networking topology, and NCCL transport internals. Quick reference for Verbs API, QP/CQ/SRQ/MR, fat-tree/rail-aligned topology, DCQCN congestion control, GPUDirect RDMA, and cluster configuration.
Navigate PyTorch torch.distributed, NCCL, and distributed training internals. Quick reference for ProcessGroupNCCL, DDP, FSDP, NCCL channels/proxy/protocol, and training configuration.
SOC 職業分類に基づく
| name | cross-framework-avoidance |
| description | Validate GRPO training configs against 36 cross-framework bug avoidance rules (19 MUST DO + 17 MUST NOT). Covers 7 frameworks (DeepSpeed, Megatron, vLLM, verl, SGLang, rLLM, MindIE), 7 pattern classes, and MoE NaN avoidance. |
| triggers | ["cross-framework bug","GRPO config validation","bug avoidance matrix","MUST DO rules","MUST NOT rules","MoE NaN","FP16 softmax","LoRA distortion","config checker","training config validation","DeepSpeed config check","Megatron config check","vLLM config check","verl config check","SGLang config check","rLLM config check","MindIE config check","36 rules","7 pattern classes"] |
| Pattern Class | DeepSpeed | Megatron | vLLM | verl | SLLM | rLLM | MindIE |
|---|---|---|---|---|---|---|---|
| cuda_stream_safety | #8061 | — | #45552 | — | #28499 | — | — |
| muon_clipping | #8068 | #5394/#5395 | — | #7776 | — | — | — |
| discrete_decision_mismatch | — | #5384/#5386 | — | — | — | — | — |
| lora_distortion | — | — | #6782 | — | #28566 | — | — |
| execution_overlap_oom | — | — | #45552 | — | — | — | #6772 |
| singleton_degeneration | — | — | — | — | — | #605/#663 | — |
| moe_fp16_nan | — | — | #10579 | — | — | — | #10579 |
All patterns marked → AVOIDED in our RTX 4090 config (except stale_cache_corruption → monitor).
| # | Rule | Why | Bug IDs |
|---|---|---|---|
| 1 | ZeRO-2 + CPU_Adam | dp=1 ZeRO-3 = pure overhead | #8061 |
| 2 | clip_grad=1.0 | Default 0.0 = silently disabled | #8068 |
| 3 | bypass_mode | Skip old_log_prob forward | verl V1 |
| 4 | group_by_prompt | Correct GRPO advantage norm | #667 |
| 5 | enforce_eager=True | Disable CUDA graph for training | safety |
| 6 | FSDP1 | Only proven safe for GRPO | #6468, #6772 |
| 7 | gs>=4 | gs=1 = REINFORCE degeneration | #605, #663 |
| 8 | LoRA r=32 + bypass sync | rank>=64 breaks EOS | #6782 |
| 9 | naive checkpoint dp=1 | No distributed overhead | perf |
| 10 | record_stream | Async CUDA stream safety | #8061, #45552 |
| 11 | ulimit 65535 | FD limit for NCCL | NCCL |
| 12 | shaped rewards | Not flat outcome rewards | #663 |
| 13 | gs>=8 for sparse MoE | Sparse needs larger groups | #605 |
| 14 | sleep_level=1 | Never sleep_level=2 RTX 4090 | risk |
| 15 | SGLang rollout | Sleep/wake support for GRPO | verl V1 |
| 16 | DSAIndexerReplay | Train/rollout consistency | #5384 |
| 17 | FP32 gating softmax | MoE NaN universal | #10579 |
| 18 | ECHO auxiliary loss | Free env-observation loss | #668 |
| 19 | LoRA+bypass sync | Delta sync, not full param | verl V1 |
| # | Rule | Why | Bug IDs |
|---|---|---|---|
| 1 | ZeRO-3 | Pure overhead dp=1 | #8061 |
| 2 | Muon optimizer | Scale-invariant + clip = stall | #5394, #8068 |
| 3 | overlap_comm=True dp=1 | NaN + no benefit | #8061 |
| 4 | sleep_level=2 RTX 4090 | Weight corruption | risk |
| 5 | LoRA rank>=64 vLLM | Breaks EOS | #6782 |
| 6 | FSDP2 | BLOCKER: CPU leak + OOM | #6468, #6772 |
| 7 | gs=1 | REINFORCE degeneration | #605, #663 |
| 8 | PPO-clip RTX 4090 | OOM 65.06 > 24 GiB | budget |
| 9 | full param sync | LoRA delta sufficient | perf |
| 10 | NCCL checkpoint dp=1 | Use naive checkpoint | perf |
| 11 | outcome reward gs<16 | Too noisy for small groups | #663 |
| 12 | merge vLLM encoder cache revert | Corruption bug | #46125 |
| 13 | global clipping Muon ChainedOptimizer | Fundamental contradiction | #5394 |
| 14 | FP16 gating softmax MoE | Universal NaN | #10579 |
| 15 | torch.compile + overlap_comm | NaN | #8061 |
| 16 | gradient_clipping=0.0 | Silently disabled | #8068 |
| 17 | bf16+fp16 both enabled | Choose bf16 only (SM89) | perf |
python3 tools/cross_framework_grpo_bug_avoidance_matrix.py matrix # 7×7 pattern matrix
python3 tools/cross_framework_grpo_bug_avoidance_matrix.py validate # Validate config (JSON)
python3 tools/cross_framework_grpo_bug_avoidance_matrix.py rtx4090 # RTX 4090 optimal config
python3 tools/cross_framework_grpo_bug_avoidance_matrix.py cross-framework # Cross-framework analysis
Validate a config:
python3 tools/cross_framework_grpo_bug_avoidance_matrix.py validate --config '{"zero_stage":2,"gradient_clipping":1.0,"overlap_comm":false,"lora_rank":32,"fsdp_mode":"fsdp1","group_size":8}'
python3 tools/moe_nan_fp32_softmax_avoidance_tool.py check # Check MoE config for NaN risks
python3 tools/moe_nan_fp32_softmax_avoidance_tool.py diagnose # Diagnose MoE NaN from symptoms
python3 tools/moe_nan_fp32_softmax_avoidance_tool.py avoid # Generate avoidance config
python3 tools/moe_nan_fp32_softmax_avoidance_tool.py rtx4090 # RTX 4090 MoE NaN report
python3 tools/megatron_muon_clipping_avoidance_tool.py check # Check Megatron config
python3 tools/megatron_muon_clipping_avoidance_tool.py diagnose # Diagnose from symptoms
python3 tools/megatron_muon_clipping_avoidance_tool.py avoid # Generate avoidance config
python3 tools/megatron_muon_clipping_avoidance_tool.py rtx4090 # RTX 4090 Muon report
python3 tools/fsdp1_vs_fsdp2_decision_guide.py guide # Decision guide
python3 tools/fsdp1_vs_fsdp2_decision_guide.py compare # Compare FSDP1 vs FSDP2
python3 tools/fsdp1_vs_fsdp2_decision_guide.py validate # Validate FSDP config
python3 tools/fsdp1_vs_fsdp2_decision_guide.py rtx4090 # RTX 4090 FSDP report
python3 tools/deepspeed_zero_safety_checker.py --mode check --config <path>
python3 tools/deepspeed_zero_safety_checker.py --mode diagnose --symptoms <str>
python3 tools/deepspeed_zero_safety_checker.py --mode rtx4090
{
"zero_stage": 2,
"optimizer": "cpu_adam",
"gradient_clipping": 1.0,
"overlap_comm": false,
"lora_rank": 32,
"fsdp_mode": "fsdp1",
"group_size": 8,
"bypass_mode": true,
"enforce_eager": true,
"use_critic": false,
"ref_in_actor": true,
"checkpoint_mode": "naive",
"sleep_level": 1,
"rollout_engine": "sglang",
"model_dtype": "bf16",
"gating_dtype": "fp32"
}
Result: 32 PASS, 0 FAIL, 2 WARN (WARN: outcome reward gs<16, stale cache corruption → monitor)
★★★★★★★★ ALWAYS compute MoE gating softmax in FP32 regardless of platform.
Mathematical proof:
FP16 max ≈ 65504
softmax_FP16(logits) = exp(x_i) / Σexp(x_j)
When any x_j > 65504 → exp(x_j) = inf → inf/inf = NaN
FP32 fix: logits.float() → softmax → result.to(dtype)
| Platform | Bug | Root Cause | Fix |
|---|---|---|---|
| CUDA (Switch TF 2021) | FP16 NaN | softmax overflow | FP32 gating |
| Ascend NPU (#10579) | torch.abs() + FP16 softmax | both patterns | Remove abs() + FP32 gating |
| Megatron (proactive) | FP32 gating built-in | Defense | Already safe |
| DeepSpeed (proactive) | FP32 gating built-in | Defense | Already safe |
RTX 4090: BF16 model dtype + FP32 gating softmax = safest config on ALL platforms.
Always load:
notebook/auto-memory.mdOn-demand load per framework:
notebook/projects/deepspeed-zero1-2-gradient-flow-stream-safety-deep-reading.mdnotebook/projects/megatron-5394-chained-optimizer-muon-clipping-reading.mdnotebook/projects/vllm-v1-architecture-critical-bugs-deep-reading.mdnotebook/projects/verl-v1-grpo-training-loop-architecture-deep-reading.mdnotebook/projects/rllm-latest-developments-2026-06-session3-reading.mdnotebook/projects/mindie-vllm-ascend-latest-developments-2026-06-session3-reading.mdnotebook/projects/pytorch-fsdp2-2026-deep-reading.md