[Intel GPU] Fix incorrect KV-cache page table for local attention when page_size > 1

[ckvermaAI]

Motivation

Fixes incorrect KV-cache page table values passed to make_local_attention_virtual_batches in the XPU (Intel GPU) attention backend when page_size > 1 and local (chunked) attention is enabled. This bug caused incorrect/zeroed outputs from flash_attn_with_kvcache.

Root Cause

make_local_attention_virtual_batches expects a page-granularity block table where each column p stores the physical page index for logical page p. However, the raw req_to_token table is token-granularity (column i = KV slot for tokeni).

When page_size > 1, the un-strided token-granularity table was passed directly, causing block_starts = k_seqstarts_absolute // page_size to index incorrect physical page values.

Modifications

When page_size > 1, the page table is first converted to page-granularity by:

1. Striding: selecting every page_size-th column (torch.arange(0, ..., page_size))
2. Dividing: integer-dividing values by page_size to convert KV slot indices to physical page indices
   This mirrors the normalization already applied to metadata.page_table elsewhere in the backend.

Changes
python/sglang/srt/layers/attention/xpu_backend.py: Add page table stride+divide normalization before passing to make_local_attention_virtual_batches when page_size > 1

Accuracy Tests

GSM8k benchmark on XPU with page_size > 1 and local attention enabled:

Metric	Before	After
Accuracy	0.005	0.815
Invalid	0.015	0.000

Speed Tests and Profiling

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.
• 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 #26011291231
Latest PR Test (Extra): ⚠️ Not enabled — add run-ci-extra label to opt in.

[gemini-code-assist]

Code Review

This pull request introduces logic to normalize the page table within the _init_local_attn_metadata function of the XPU backend, ensuring it is at page-granularity when page_size is greater than one. Feedback suggests refactoring this logic into a shared helper method to eliminate duplication with similar code found in init_forward_metadata.

[gemini-code-assist]

The normalization logic added here (striding and floor-dividing the page table) is identical to the logic at lines 371-376 in init_forward_metadata. While necessary here because _init_local_attn_metadata is called before that main normalization block, it would be cleaner to encapsulate this logic into a helper method or move the main normalization earlier in init_forward_metadata to avoid duplication and ensure consistency across the backend.

[mingfeima]

@sunjiweiswift please help review this one!

[gemini-code-assist]

Warning

You have reached your daily quota limit. Please wait up to 24 hours and I will start processing your requests again!

[mingfeima]

we'd better add UT to cover radix attention level behavior, like this one:

• test/registered/attention/test_triton_attention_backend.py
• test/registered/attention/test_torch_native_attention_backend.py
• test/registered/cpu/test_intel_amx_attention_backend_a.py

@ckvermaAI could you please take this job?

[ckvermaAI]

@mingfeima sure, Let me add similar unit test

[sunjiweiswift]

Let me test this on more models.

[mingfeima]

Let me test this on more models.

@sunjiweiswift any updates?