Adapted to pytest framework

Menglths · Menglths · commit 55e817f1deec · 2025-11-14T09:27:40.000+08:00
diff --git a/test/common/llmperf/run_inference.py b/test/common/llmperf/run_inference.py
@@ -95,6 +95,7 @@ def run_test_cases(test_cases, timestamp_dir, model, server_url, tokenizer_path)
                     tokenizer_path=tokenizer_path,
                     user_metadata={"case_idx": i, "phase": "prefill"},
                 )
+                reset_prefill_cache(env, server_url)
                 # Then run normal mode
                 print("[INFO] Prefill completed, switching to normal mode execution")
                 summary = run_token_benchmark(

Original file line number	Diff line number	Diff line change
`@@ -95,6 +95,7 @@ def run_test_cases(test_cases, timestamp_dir, model, server_url, tokenizer_path)`
`95`	`95`	`tokenizer_path=tokenizer_path,`
`96`	`96`	`user_metadata={"case_idx": i, "phase": "prefill"},`
`97`	`97`	`)`
	`98`	`+ reset_prefill_cache(env, server_url)`
`98`	`99`	`# Then run normal mode`
`99`	`100`	`print("[INFO] Prefill completed, switching to normal mode execution")`
`100`	`101`	`summary = run_token_benchmark(`