A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 5, 04:45 UTC
[CI][KVConnector][Metrics] Update multi KV connector edge case according to prefill stats changes (#39808)
Signed-off-by: Zhanqiu Hu <zhu@redhat.com>
zhanqiuhu committed
Apr 14, 2026 at 14:59 UTC
30679319e83a78f6fd405130e562043c1e7f5e7e
1 file changed
+11
-11
tests/v1/kv_connector/nixl_integration/test_multi_connector_edge_cases.py
+11
-11
@@ -297,7 +297,7 @@ def test_multi_block_correctness():
297
298
299
def test_cold_decode_no_cache_hit_metrics():
300
- """Cold decode: external_kv_transfer==P, local_cache_hit==0."""
300
+ """Cold decode: external_kv_transfer==P, local_cache_hit==0, local_compute==0."""
301
n0 = _fetch_nixl_bytes(DECODE_HOST, DECODE_PORT)
302
m0 = _fetch_decode_metrics()
303
proxy_text, P = _complete(proxy_client, MEDIUM_PROMPT)
@@ -312,8 +312,8 @@ def test_cold_decode_no_cache_hit_metrics():
312
assert d["external_kv_transfer"] == P, (
313
f"expected external_kv_transfer={P}, got {d['external_kv_transfer']}"
314
)
315
- assert d["local_compute"] == 1, (
316
- f"expected local_compute=1, got {d['local_compute']}"
315
+ assert d["local_compute"] == 0, (
316
+ f"expected local_compute=0, got {d['local_compute']}"
317
)
318
assert d["local_cache_hit"] == 0, (
319
f"expected local_cache_hit=0, got {d['local_cache_hit']}"
@@ -341,15 +341,15 @@ def test_full_decode_gpu_cache_hit_metrics():
341
print(f"FULL CACHE HIT: {P} tokens, cached={cached}, nixl={expected_nixl}")
342
print(f" metrics delta: {d}, nixl_bytes_delta={n1 - n0}")
343
assert len(proxy_text) > 0, "proxy returned empty response"
344
- assert d["local_cache_hit"] == cached - 1, (
345
- f"expected local_cache_hit={cached - 1}, got {d['local_cache_hit']}"
344
+ assert d["local_cache_hit"] == cached, (
345
+ f"expected local_cache_hit={cached}, got {d['local_cache_hit']}"
346
)
347
assert d["external_kv_transfer"] == expected_nixl, (
348
f"expected external_kv_transfer={expected_nixl}, "
349
f"got {d['external_kv_transfer']}"
350
)
351
- assert d["local_compute"] == 1, (
352
- f"expected local_compute=1 (recomputed last token), got {d['local_compute']}"
351
+ assert d["local_compute"] == 0, (
352
+ f"expected local_compute=0, got {d['local_compute']}"
353
)
354
assert n1 - n0 > 0, (
355
f"expected nixl_bytes_transferred to increase (partial NIXL for "
@@ -383,11 +383,11 @@ def test_partial_decode_gpu_cache_hit_metrics():
383
f"expected external_kv_transfer={expected_nixl}, "
384
f"got {d['external_kv_transfer']}"
385
)
386
- assert d["local_cache_hit"] == cached - 1, (
387
- f"expected local_cache_hit={cached - 1}, got {d['local_cache_hit']}"
386
+ assert d["local_cache_hit"] == cached, (
387
+ f"expected local_cache_hit={cached}, got {d['local_cache_hit']}"
388
)
389
- assert d["local_compute"] == 1, (
390
- f"expected local_compute=1 (recomputed last token), got {d['local_compute']}"
389
+ assert d["local_compute"] == 0, (
390
+ f"expected local_compute=0, got {d['local_compute']}"
391
)
392
assert n1 - n0 > 0, (
393
f"expected nixl_bytes_transferred to increase (NIXL for uncached "