Complete opt-in Qwen HiCache state transfer and backport QSA/restore fixes

[ktsaou]

Qwen HiCache previously restored KV/Mamba buffers without all PLE and compressed QSA state, so a reused prefix could differ from its original computation. This updates the isolated HiCache profile onto current main, carries the missing target/draft state through RAM and files, and adds upstream QSA, router and restore-order corrections.

The ordinary Dockerfile and default profiles remain unchanged. The opt-in build pins the cumulative parent and verifies the complete 4,393-file runtime inventory.

Changes:

• PLE convolution/N-gram checkpoints, host sizing and layer waits.
• Compressed QSA target/draft sidecar, complete-page requirements and file-miss handling.

Upstream backports/adaptations:

Upstream SGLang PR	Local patch	Change used
#38851	0023	Initialize page-strided QSA scratch, use 64-bit gather offsets, and convert FP8 gathered values correctly.
#38290	0024	Wait before reading router bias in both Triton and radix kernels. The existing bias API is retained.
#36738	0025	Order H2D restoration after in-flight forwards that may still write reused pages.
#36743	0025	Wait for the final restored layer before copying an entire Mamba slot.
#39446	0026	Clamp compressed-key gather rows for short cached extensions.

These changes were adapted to the pinned cumulative runtime. PRs #36743 and #39446 were open when imported; the included regression tests validate the applied changes locally. Patches 0020–0022 contain the Qwen companion-state work preserved from this draft.

Validation:

• 67 cache CPU tests, 95 repository tests, and focused GPU gather/router/restore checks.
• Post-restoration tools 48/48, mixed efforts 96/96, and image-bearing tool results 16/16.
• A 603-second mixed-load run reached 64 active requests: media 32/32, exact 523,264-token retrieval, tools 288/288, and clean cancellation/drain without background client errors.
• Disk reuse across a full restart returned 4/4 correct answers with explicit disk hits and measured H2D restoration; final tool replay passed 48/48.
• Short-prompt decode measured 249 tokens/s at C1 and 2,843 aggregate tokens/s at C64. This is not a matched cache-on/off comparison.

The tested TP2/MTP3 configuration retains scheduler overlap, target/draft graphs, 64 request slots, 512k context and multimodal inputs, using 60GB aggregate RAM HiCache and a 512GB file payload cap. Prefill CUDA graphs remain disabled as in the baseline; no additional performance feature was disabled.

Limits: only the kernel transfer path and built-in file backend are covered. Best-effort prefetch can recompute, so the strict all-request disk-hit assertion remains recorded as failed; correct recomputation alone is not counted as restoration. One prefetched request reported a device hit. RAM-only pressure tests did not establish host hits, while the combined disk path exercised RAM-to-GPU restoration. Very short video ordering remains a known limitation; longer multi-frame fixtures pass.

Configuration, historical failures and detailed results: docs/hicache-wip.md.