[hnyls2002]

• The prefill CUDA graph's captured full top-k (FullTopKIndexer._deep_gemm_prefill_captured) now selects each row up to its own length with topk_transform_paged_from_metadata, the selection the decode path already runs on the same paged logits, instead of a torch.topk over the whole captured width.
• The v1 kernel takes k from the output width and caps it at 1024 (kMaxTopK in deepseek_v4/topk_v1.cuh, now TOPK_V1_MAX_K). Outputs wider than that keep the torch top-k over min(k, width) columns, so no configured k is truncated.
• TestDSV4FullTopKCapturedFallback checks which path each output width takes, and that the torch path keeps every column.

Behavior change: with --cuda-graph-backend-prefill breakable, DeepSeek-V4.1's ratio-1/2 indexer layers inside the prefill graph select with the per-row kernel (index_topk is 512 for DeepSeek-V4.1-Flash).

Checked on GB300 with DeepSeek-V4.1-Flash, TP4 (DSPARK, chunked prefill 8192, decoder replay on), against main (617c8b6) with the same prefill graph:

• test_deepseek_v4.py passes on GB300 (27 tests).
• GSM8K (sgl-eval, 200 questions, the model's preset, 8-shot prefix): 98.0%, against main's 97.5%.
• First-token top-20 distributions on prompts that replay the graph (590 to 14,795 tokens), as total variation distance from main eager: at most 0.073; main's graph is at 0.133 on the same prompt, and a second main eager run at 0.088.
• TTFT, one request, median of 3:

New tokens	main (ms)	this PR (ms)
256	34.6	34.7
1,024	44.9	41.8
8,000	109.8	101.1

---

CI States

Latest PR Test (Base): ❌ Run #39192645755
Latest PR Test (Extra): ❌ Run #39192645634
Latest PR Test (AMD ROCm 10): ❌ Run #39192645601

[hnyls2002]

Closing: #45783 drops the captured full top-k this PR changed. Since #43733 the released V4.1 checkpoints run the two-level path in the prefill graph, so this code was reached only through the torch prefill indexer switch or a V4.1 config without candidate blocks.