Revert "[nvbug 5304752][fix] enhance _check_arguments to filter illegal requests for pytorch backend (NVIDIA#5541)"

This reverts commit 388b491.

Signed-off-by: Pengyun Lin <[email protected]>

Author: LinPoly

tensorrt_llm/llmapi/llm.py

@@ -529,13 +529,6 @@ def _check_arguments(self, prompt_len: int, query_len: int,
                 raise ValueError(
                     f"PyTorch backend currently only supports `logprobs=1`. Received `logprobs={sampling_params.logprobs}` (Top{sampling_params.logprobs} logprobs). Please set `logprobs=1` in `sampling_params` instead."
                 )
-            # Check prompt length and query length against max_num_tokens to filter illegal requests.
-            if self.args.backend == "pytorch" and not self.args.enable_chunked_prefill:
-                max_num_tokens = self.args.max_num_tokens
-                if max_num_tokens and prompt_len / self.args.parallel_config.cp_size + query_len > max_num_tokens:
-                    raise ValueError(
-                        f"The sum of prompt length ({prompt_len/self.args.parallel_config.cp_size}), query length ({query_len}) and max_tokens ({sampling_params.max_tokens}) should not exceed "
-                        f"max_num_tokens ({max_num_tokens})")
             return
 
         build_config = self.args.build_config
@@ -552,7 +545,7 @@ def _check_arguments(self, prompt_len: int, query_len: int,
             (sampling_params.max_tokens or 0) > max_seq_len):
             raise ValueError(
                 f"The sum of prompt length ({prompt_len/self.args.parallel_config.cp_size}) and query length ({query_len}) max_tokens ({sampling_params.max_tokens}) should not exceed "
-                f"max_seq_len ({max_seq_len})")
+                f"max_seq_len ({build_config.max_seq_len})")
 
         if sampling_params.use_beam_search and sampling_params.best_of > build_config.max_beam_width:
             if sampling_params.n == sampling_params.best_of:

tests/unittest/llmapi/test_llm.py

@@ -2060,37 +2060,24 @@ def success_path():
     success_path()
 
 
-def _test_llm_capture_request_error(pytorch_backend: bool, tp_size: int = 1):
-    llm_args_extra = {}
-    if pytorch_backend:
-        from tensorrt_llm._torch import LLM as LLM_torch
-        LLM_CLASS = LLM_torch
-        llm_args_extra["max_num_tokens"] = 64
-    else:
-        LLM_CLASS = LLM
-        build_config = BuildConfig()
-        build_config.max_num_tokens = 64
-        llm_args_extra["fast_build"] = True
-        llm_args_extra["build_config"] = build_config
+def _test_llm_capture_request_error(tp_size: int = 1):
+    build_config = BuildConfig()
+    build_config.max_num_tokens = 64
 
-    llm = LLM_CLASS(
+    llm = LLM(
         model=llama_model_path,
-        tensor_parallel_size=tp_size,
-        **llm_args_extra,
+        build_config=build_config,
+        fast_build=True,
     )
 
     prompt = 'A ' * 65  # the minimum max_num_tokens is 64
-    if pytorch_backend:
-        # pytorch backend will raise ValueError for max_num_tokens
-        with pytest.raises(ValueError):
-            llm.generate(prompt)
-    else:
-        with pytest.raises(RequestError):
-            llm.generate(prompt)
+
+    with pytest.raises(RequestError):
+        llm.generate(prompt)
 
 
 def test_llm_capture_request_error():
-    _test_llm_capture_request_error(pytorch_backend=False, tp_size=1)
+    _test_llm_capture_request_error(tp_size=1)
 
 
 def test_llm_api_jupyter_scenario():

tests/unittest/llmapi/test_llm_multi_gpu.py

@@ -451,7 +451,7 @@ def test_llm_get_stats_async_tp2(pytorch_backend):
 
 
 def test_llm_capture_request_error():
-    _test_llm_capture_request_error(pytorch_backend=False, tp_size=2)
+    _test_llm_capture_request_error(tp_size=2)
 
 
 def test_llm_with_postprocess_parallel_tp2():

tests/unittest/llmapi/test_llm_multi_gpu_pytorch.py

@@ -5,17 +5,11 @@
 from tensorrt_llm.llmapi import KvCacheConfig
 from .test_llm_pytorch import (llama_v2_13b_lora_test_harness,
                                llama_7b_multi_lora_test_harness)
-from .test_llm import _test_llm_capture_request_error
 # isort: on
 
 global_kv_cache_config = KvCacheConfig(free_gpu_memory_fraction=0.4)
 
 
-@pytest.mark.gpu2
-def test_llm_capture_request_error():
-    _test_llm_capture_request_error(pytorch_backend=True, tp_size=2)
-
-
 @pytest.mark.gpu4
 def test_tinyllama_logits_processor_tp2pp2():
     tinyllama_logits_processor_test_harness(backend="pytorch",

tests/unittest/llmapi/test_llm_pytorch.py

@@ -4,11 +4,12 @@
 from tensorrt_llm.sampling_params import SamplingParams
 
 # isort: off
-from .test_llm import (
-    get_model_path, global_kvcache_config, llama_model_path,
-    llm_get_stats_async_test_harness, llm_get_stats_test_harness, prompts,
-    run_llm_abort_request, run_llm_with_postprocess_parallel_and_result_handler,
-    tinyllama_logits_processor_test_harness, _test_llm_capture_request_error)
+from .test_llm import (get_model_path, global_kvcache_config, llama_model_path,
+                       llm_get_stats_async_test_harness,
+                       llm_get_stats_test_harness, prompts,
+                       run_llm_abort_request,
+                       run_llm_with_postprocess_parallel_and_result_handler,
+                       tinyllama_logits_processor_test_harness)
 from utils.util import force_ampere, similar, skip_gpu_memory_less_than_40gb, skip_gpu_memory_less_than_80gb, skip_gpu_memory_less_than_138gb
 from utils.llm_data import llm_models_root
 from tensorrt_llm.lora_manager import LoraConfig
@@ -63,10 +64,6 @@ def test_llm_get_stats_async(return_context_logits, use_overlap,
         enable_iter_req_stats=enable_iter_req_stats)
 
 
-def test_llm_capture_request_error():
-    _test_llm_capture_request_error(pytorch_backend=True, tp_size=1)
-
-
 @force_ampere
 @pytest.mark.parametrize(
     "sampling_params",