[Bugfix] Limit profiling run sequence length by max_model_len (#14785)

Signed-off-by: Kyle Sayers <kylesayrs@gmail.com>

vllm/inputs/registry.py

@@ -330,6 +330,11 @@ class InputRegistry:
         from vllm.multimodal import MultiModalKwargs
         from vllm.multimodal.profiling import MultiModalProfiler
 
+        if seq_len > model_config.max_model_len:
+            raise AssertionError(
+                f"Profiling attempted with sequence length ({seq_len}) "
+                f"greater than model length ({model_config.max_model_len})")
+
         if mm_registry.has_processor(model_config):
             tokenizer = cached_tokenizer_from_config(model_config)
             processor = mm_registry.create_processor(model_config,

vllm/worker/enc_dec_model_runner.py

@@ -281,6 +281,7 @@ class EncoderDecoderModelRunner(GPUModelRunnerBase[EncoderDecoderModelInput]):
         for group_id in range(max_num_seqs):
             seq_len = (max_num_batched_tokens // max_num_seqs +
                        (group_id < max_num_batched_tokens % max_num_seqs))
+            seq_len = min(seq_len, self.model_config.max_model_len)
             batch_size += seq_len
 
             decoder_dummy_data = self.input_registry \

vllm/worker/model_runner.py

@@ -1302,6 +1302,7 @@ class GPUModelRunnerBase(ModelRunnerBase[TModelInputForGPU]):
             for group_id in range(max_num_seqs):
                 seq_len = (max_num_batched_tokens // max_num_seqs +
                            (group_id < max_num_batched_tokens % max_num_seqs))
+                seq_len = min(seq_len, self.model_config.max_model_len)
                 batch_size += seq_len
 
                 dummy_data = self.input_registry \

vllm/worker/openvino_model_runner.py

@@ -148,6 +148,7 @@ class OpenVINOModelRunner(ModelRunnerBase):
                 seq_len = min(
                     seq_data.get_len(),
                     computed_len + seq_group_metadata.token_chunk_size,
+                    self.model_config.max_model_len,
                 )
                 if is_prompt:
                     tokens = seq_data.get_token_ids()[computed_len:seq_len]

vllm/worker/xpu_model_runner.py

@@ -466,6 +466,7 @@ class XPUModelRunner(ModelRunnerBase[ModelInputForXPUWithSamplingMetadata]):
         for group_id in range(max_num_seqs):
             seq_len = (max_num_batched_tokens // max_num_seqs +
                        (group_id < max_num_batched_tokens % max_num_seqs))
+            seq_len = min(seq_len, self.model_config.max_model_len)
             batch_size += seq_len
 
             dummy_data = self.input_registry \