[gspark-etri]

Background

AtomKVBlocks accepts kv_blocks == lanes + 1 (one block per lane plus the scratch block). With prefix caching on, a finished request's block becomes the only cache entry; when a new request in that lane shares its prefix, admit() calls _take(keep=src), finds no victim other than the source, raises RuntimeError("no KV block available"), and step() poisons the batcher. Repro: batch=1, blocks=2, the same prompt twice.

Changes

• opt_prime/opt_prime/accel/platforms/atom/batching.py:81-82: with prefix_cache, AtomKVBlocks requires kv_blocks >= lanes + 2 and raises ValueError otherwise. When a lane is admitted at most lanes - 1 blocks are held by lanes, so lanes + 2 blocks always leave an entry other than the copy source (or a free block). The AtomBatcher docstring states the requirement.
• opt_prime/tests/test_atom_batching.py: test_prefix_cache_needs_an_evictable_block checks the rejection, and that kv_blocks == lanes + 1 without prefix cache still serves the same prompt twice token-exactly.

I chose rejection over reusing the source block in place: it is a two-line change, and with a single cache block the reuse path would only help the exact-repeat case at the cost of extra special cases in _take/admit (skipping the copy, dropping the index entries, LRU bookkeeping).

Verification

CPU only (fake RBLN SDK of the existing tests, no NPU), Python 3.10, torch 2.10.0+cu128, base optimus-rm @253be20.

On the unfixed code (from opt_prime/):

TORCH_DEVICE_BACKEND_AUTOLOAD=0 CUDA_VISIBLE_DEVICES=1 PYTHONPATH=$PWD python3 -m pytest -q tests/test_atom_batching.py -k evictable
E  Failed: DID NOT RAISE ValueError
1 failed

and the issue's repro (same prompt twice, batch=1, blocks=2, prefix cache) fails with RuntimeError: no KV block available at batching.py:172. After the fix the new test passes; tests/test_atom_batching.py tests/test_pd_atom_reuse.py tests/test_pd_native_atom.py: 15 passed, 2 skipped.

Whole suite:

TORCH_DEVICE_BACKEND_AUTOLOAD=0 CUDA_VISIBLE_DEVICES=1 PYTHONPATH=$PWD python3 -m pytest -q tests
before (253be20): 400 passed, 23 skipped, 135 subtests passed
after:            401 passed, 23 skipped, 135 subtests passed

Compatibility and limits

• Configurations with prefix_cache=True and kv_blocks == batch_size + 1 now fail at construction instead of failing on the first prefix hit; they need one more compiled KV block. docs/validation/batching/ATOM.md already describes kv_blocks = B + cache + 1.
• Own block tables without prefix cache (PD import, plain refill) still accept kv_blocks == lanes + 1.

Related Issue

Closes #172

🤖 Generated with Claude Code