fix: profiling script missing tests when kv cache is tight (#2567)

benchmarks/profiler/utils/profile_decode.py

@@ -42,13 +42,20 @@ def profile_decode(
         (max_context_length - osl) // interpolation_granularity,
     ):
         max_concurrency = max_kv_tokens // (isl + osl)
-        if max_concurrency // interpolation_granularity == 0:
+        if max_concurrency == 0:
+            logger.warning(
+                f"max_kv_tokens {max_kv_tokens} is too small for"
+                f" isl {isl} + osl {osl}, skipping."
+            )
+            break
+        elif max_concurrency < interpolation_granularity:
             logger.warning(
                 f"max_concurrency {max_concurrency} is too small for"
                 f" interpolation granularity {interpolation_granularity}."
                 f" max_kv_tokens {max_kv_tokens}, isl {isl}, osl {osl}"
             )
-            break
+            sweep_num_request = range(1, max_concurrency + 1)
+        else:
             sweep_num_request = range(
                 1,
                 max_concurrency,