[YAMY1234]

Motivation

Breakable CUDA Graph pads token buffers to a captured bucket. The attention wrapper already narrows Q/K/V and cache locations to the real token count, but it still passes padded positions to the backend. Decode context parallelism (DCP) uses these positions to determine key-value (KV) ownership, so virtual padded tokens can compete with real tokens for the same physical KV slots and silently corrupt the cache.

Modifications

• Narrow forward_batch.positions to the real query-token count together with out_cache_loc for the attention call.
• Restore both tensors on the shared ForwardBatch after the call.
• Extend the existing Breakable CUDA Graph test to verify the narrowed backend view and restored batch metadata.

Accuracy Tests

The targeted GPU regression was run before and after the change:

• Before: the attention backend received five padded positions for three real cache locations, and the test failed.
• After: the backend received only the three real positions. The targeted test passed, and the full TestBreakableCUDAGraphBasic suite passed (9 passed).

An end-to-end matched A/B used Qwen3.5-397B-A17B FP8 with TP16/DCP4, FP8 KV cache, Breakable CUDA Graph prefill, full decode CUDA Graph, GSM8K-128, 8-shot prompting, temperature 0, and max_tokens=16384. The validation stack included #32795 and #32858; the only runtime-code difference between before and after was this PR's four-line positions fix.

Version	Runs	GSM8K
Before	1	0.875 (112/128)
After	3	0.984375 (126/128) in every run

All four runs completed 128/128 requests with HTTP 200 and no non-200 responses.

Speed Tests and Profiling

Not benchmarked. This correctness fix does not change kernels or CUDA Graph configuration; it creates a tensor view and restores the original reference around the existing attention call.

Checklist

• Format your code according to the Format code with pre-commit.
• Add unit tests according to the Run and add unit tests.
• Update documentation according to Write documentations. No user-facing documentation change is required.
• Provide accuracy and speed benchmark results according to Test the accuracy and Benchmark the speed.
• Follow the SGLang code style guidance.

Review and Merge Process

1. Ping Merge Oncalls to start the process. See the PR Merge Process.
2. Get approvals from CODEOWNERS and other reviewers.
3. Trigger CI tests with comments or contact authorized users to do so.
   • Common commands include /tag-and-rerun-ci, /tag-run-ci-label, /rerun-failed-ci
4. After green CI and required approvals, ask Merge Oncalls or people with Write permission to merge the PR.

---

CI States

Latest PR Test (Base): ✅ Run #30979497752
Latest PR Test (Extra): ❌ Run #30979497576

[gemini-code-assist]

Caution

The consumer version of Gemini Code Assist on GitHub has been sunset. All code review activity has officially ceased.

[gemini-code-assist]

Caution

The consumer version of Gemini Code Assist on GitHub has been sunset. All code review activity has officially ceased.

[Oasis-Git]

Wait CI for merge

[YAMY1234]

/tag-and-rerun-ci

[YAMY1234]

TestBreakableCUDAGraphBasic and all NVIDIA CI tests pass.