[Bug]: Limit num_reqs in dummy_run when max_num_seqs is small (vllm-project#26144)

Signed-off-by: Benjamin Chislett <[email protected]>

Author: benchislett

vllm/v1/worker/gpu_model_runner.py

@@ -3060,7 +3060,7 @@ def _dummy_run(
             assert not uniform_decode
             # Create mixed batch:
             # first half decode tokens, second half one prefill
-            num_decode_tokens = num_tokens // 2
+            num_decode_tokens = min(max_num_reqs - 1, num_tokens // 2)
             num_prefill_tokens = num_tokens - num_decode_tokens
             num_reqs = num_decode_tokens + 1
 
@@ -3072,7 +3072,7 @@ def _dummy_run(
             max_query_len = num_prefill_tokens
         elif uniform_decode:
             assert not create_mixed_batch
-            num_reqs = cdiv(num_tokens, max_query_len)
+            num_reqs = min(max_num_reqs, cdiv(num_tokens, max_query_len))
             num_scheduled_tokens_list = [max_query_len] * num_reqs
             if num_tokens % max_query_len != 0:
                 num_scheduled_tokens_list[-1] = num_tokens % max_query_len