docs(rocm): publish G10/G11 200-step RL results

[inaniloquentee]

What this PR adds

• Publishes the completed ROCm Qwen3-8B 200-step G10 production P/P and G11 RL-Kernel R/R consistency experiment derived from PR perf(rocm): MFMA batch-invariant GEMM and chunked Triton attention for strict R/R #396.
• Adds the complete paired-step CSV, structured summary, strict-validation summary, and two reproducible consistency figures.
• Adds local-log regeneration and W&B upload scripts so the complete RL history and source artifacts can be reproduced and published without rerunning training.

Completed experiment scope

Group	Attention / FFN / logp route	Ray submission	Result
G10	production / production (P/P)	raysubmit_Rj3ZWxq3mphM2bFL	SUCCEEDED, 200/200
G11	RL-Kernel / RL-Kernel (R/R), strict consistency	raysubmit_twpgQ7pxvniaXWXC	SUCCEEDED, 200/200

Both runs used the same frozen workload and source configuration. Only the intended G10 P/P versus G11 R/R implementation routes and identity-derived paths differ.

Item	Matched value
Hardware	1 node, 8x AMD Instinct MI300X 192GB
Actor parallelism	TP4 / CP2 / PP1
Rollout engines	2 engines, TP4 each
Horizon	200 rollout/training steps
Seeds	training 1234, rollout 1234
Sampling	1 prompt x 8 samples = global batch 8
Response limit	7,168 tokens
Dynamic batching	maximum 4,096 tokens/GPU
vLLM memory utilization	0.38
Rollout log probabilities	enabled
Reference model / KL loss	enabled, coefficient 0.001
Torch profiler	disabled

Version provenance

Arm	RL-Kernel	VIME	Megatron-LM	Notes
G10	7a9f3b5	c80200e	1dcf0da	Native P/P route
G11	7a9f3b5	c80200e	1dcf0da	Strict RL-Kernel R/R route
Result bundle	86877e4	-	-	CSV, JSON, Markdown, two consistency plots, and upload tooling

Strict train/rollout consistency and reward

• G11 passes strict validation across all 200 steps: mismatch_count == 0, max_abs_diff == 0, and torch.equal == true over 9,400,614 compared elements from 1,600 samples.
• G10 completed all 200 steps. Its final validator only rejects missing RL-Kernel operator readbacks, which the native P/P route does not emit; the training result itself is complete.
• Frozen-input and frozen-source audits match across both arms.

Mean train/rollout logp difference

Data integrity and method

• Both Ray submissions completed successfully with 200/200 recorded rollout and training steps.
• The paired CSV contains every available scalar RL, training, consistency, and timing field from both launcher logs. Required fields are finite; no imputation or row deletion is used.
• The structured summary preserves the full 200-step arithmetic view, a rollout-0-warmup-excluded token-normalized view, and the equal-length matched-step cross-check.
• The raw launcher logs and full validation JSON are included by the W&B uploader as source artifacts.

Reproducible artifacts

• Result README
• All 200 paired steps
• Distribution and bootstrap summary
• Strict validation summary
• Figure regeneration script
• W&B upload script

Validation

• Pre-commit formatting, import, and lint checks pass for both Python scripts.
• A clean regeneration from the two authoritative launcher logs produces only the two consistency PNG files by default.
• Performance plots are generated only with --include-performance-plots for external W&B reporting.
• The W&B uploader completed an offline end-to-end dry run for G10, G11, and the combined report artifact.

[coderabbitai]

Important

Review skipped

Auto reviews are disabled on base/target branches other than the default branch.

Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: defaults

Review profile: CHILL

Plan: Advanced

Run ID: 45c4ac47-d351-48fb-a224-28db1c083821

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

---

_{Comment @coderabbitai help to get the list of available commands.}

[Flink-ddd]

LGTM

[maxiaosong1124]

LGTM!