=== §3 fi-sampling enablement evidence snapshot (card7, port 8441) ===
date: 2026-09-11T15:44:06Z

--- /health ---
HTTP 200

--- extension startup / gating proof (verbatim) ---
26:INFO 09-11 15:41:03 [cascade_plugin.py:147] cascade plugin loaded (gate=1, graph_gate=1, kernel_wheel=ok)
59:(EngineCore pid=1199290) INFO 09-11 15:41:32 [cascade_plugin.py:147] cascade plugin loaded (gate=1, graph_gate=1, kernel_wheel=ok)
60:(EngineCore pid=1199290) fi_sampling sampling path is ACTIVE (VLLM_HUST_FI_SAMPLING=1): untruncated sampling uses the flashinfer-semantics triton-ascend kernel, joint top-k/top-p at B>=256 with k>=32 uses the joint kernel and k=1 rows use argmax; every other case falls back to the fork chain. Output distribution matches flashinfer semantics, not the fork's exponential-race chain (see docs/fi-sampling.md).
106:(EngineCore pid=1199290) 
Capturing CUDA graphs (mixed prefill-decode, FULL):   0%|          | 0/3 [00:00<?, ?it/s]2026-09-11 15:42:47,616 - WARNING - FIA decode de-mask applied: dropped torch.Size([2048, 2048]) int8 atten_mask and set sparse_mode 3->0 on a pure-decode FIA call (TND, Q_S=1 per request, pre/next=INT_MAX).
156:(APIServer pid=1198388) INFO:     Application startup complete.

--- fi-sampling route evidence (verbatim) ---
(EngineCore pid=1199290) 2026-09-11 15:43:15,652 - WARNING - [fi-sampling trace] route=fi_api1 B=1 k=None..None top_k=False top_p=False generators=0 logprobs_mode=raw_logprobs
(EngineCore pid=1199290) 2026-09-11 15:43:15,703 - WARNING - [fi-sampling trace] route=fi_api1 B=1 k=None..None top_k=False top_p=False generators=0 logprobs_mode=raw_logprobs
(EngineCore pid=1199290) [fi-sampling histogram calls=75 max_B=1] {'fi_api1': 75}

--- host-drift warning (expected, documented) ---
(EngineCore pid=1199290) 2026-09-11 15:43:15,652 - WARNING - fi_sampling: this vllm-ascend host does not define the ascend-config knob 'enable_async_exponential'; treating the feature it gates as disabled (feature removed upstream).

--- health counters ---
fi_sampling sampling path is ACTIVE    1
cascade plugin loaded                  2
de-mask applied                        1
route=fi_api1                          20
route=fork                             0
TypeError                              0
Traceback                              0
OOM                                    0
OutOfMemory                            0
capture body SUCCESS                   0
cascade graph capture failed           0

--- chat smoke #1 (untruncated) ---
{"id":"chatcmpl-8f022824f21fea44","object":"chat.completion","created":1789141395,"model":"/data/shared_models/Qwen--Qwen2.5-14B-Instruct","choices":[{"index":0,"message":{"role":"assistant","content":"hello fi-sampling","refusal":null,"annotations":null,"audio":null,"function_call":null,"reasoning":null},"logprobs":null,"finish_reason":"stop","stop_reason":null,"token_ids":null,"routed_experts":null}],"service_tier":null,"system_fingerprint":"vllm-0.28.1.post1.dev143+gf18cf803c-52c96ef0","usage":{"prompt_tokens":37,"total_tokens":42,"completion_tokens":5,"prompt_tokens_details":null,"completion_tokens_details":null},"prompt_logprobs":null,"prompt_token_ids":null,"prompt_text":null,"kv_transfer_params":null,"ec_transfer_params":null,"metrics":null}
HTTP_CODE=200

--- chat smoke #2 (80-token untruncated; drives the histogram) ---
{"id":"chatcmpl-9a22ddefd0d4a81d","object":"chat.completion","created":1789141410,"model":"/data/shared_models/Qwen--Qwen2.5-14B-Instruct","choices":[{"index":0,"message":{"role":"assistant","content":"1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 2","refusal":null,"annotations":null,"audio":null,"function_call":null,"reasoning":null},"logprobs":null,"finish_reason":"length","stop_reason":null,"token_ids":null,"routed_experts":null}],"service_tier":null,"system_fingerprint":"vllm-0.28.1.post1.dev143+gf18cf803c-52c96ef0","usage":{"prompt_tokens":44,"total_tokens":124,"completion_tokens":80,"prompt_tokens_details":null,"completion_tokens_details":null},"prompt_logprobs":null,"prompt_token_ids":null,"prompt_text":null,"kv_transfer_params":null,"ec_transfer_params":null,"metrics":null}
--- npu-smi card7 ---
	HBM Usage Rate(%)              : 86
