[None][fix] kv_cache_v2: fix SWA scratch on deferred-context revert under ADP balance

[qiaoxj07]

Summary

revert_allocate_context() undoes the per-iteration resize_context() capacity grow when attention-DP balance (attention_dp_config.enable_balance) or batch-waiting defers a context request after V2 scheduling. On sliding-window-attention (SWA) layers this collided with SWA scratch reuse and crashed DeepSeek-V4 DEP8 runs once adp_balance was enabled.

Two failure modes:

1. A freshly-deferred context (history_length == 0) tripped resize()'s SWA-scratch invariant history_length >= old_capacity - max_rewind_len (which only holds for a generation-time window slide) — e.g. AssertionError: SWA scratch requires old_capacity - max_rewind_len (965) <= history_length (0) <= old_capacity (965).
2. A context whose prefill grew past the sliding window holds scratch slots, and resize() refuses to shrink while scratch slots are live.

Fix

• _KVCache.resize() gains a keyword-only allow_swa_scratch_reuse (default True). The revert path passes False to perform a scratch-free capacity shrink, leaving the generation-time window-slide path untouched.
• revert_allocate_context() suspend()s caches that hold scratch slots — freeing the scratch slots and releasing pages back to the pool — and resumes them via prepare_context() / _resume_and_restore() on the next schedule (the same machinery generation and the existing pre_cap > 0 revert already use).

adp_balance stays fully functional and deferred pages are still freed during the wait window (no KV-cache capacity regression).

Test

DSv4 DEP8 (TP8/EP8, attention-DP), adp_balance (timeout_iters=100, batching_wait_iters=20), SWA, MTP0, NVFP4, bs128 / conc256 on GB200:

• Before: crashes in warmup with SWA scratch requires ....
• After: the full benchmark completes (Total Token Throughput ~13.6k tok/s) with 0 SWA assertions and 0 errors.

🤖 Generated with Claude Code

[jiaganc]

This issue is fixed in #14849. I suggest you to use that PR.