Improve benchmark scripts & fix llava (#613)

benchmark/latency_throughput/README.md

@@ -30,7 +30,7 @@ wget https://huggingface.co/datasets/anon8231489123/ShareGPT_Vicuna_unfiltered/r
 
 #### Run ShareGPT
 ```
-python3 bench_throughput.py --backend srt --port 30000 --tokenizer meta-llama/Llama-2-7b-chat-hf --dataset ShareGPT_V3_unfiltered_cleaned_split.json --num-prompts 10 --request-rate 10
+python3 bench_serving.py --backend srt --port 30000 --tokenizer meta-llama/Llama-2-7b-chat-hf --dataset ShareGPT_V3_unfiltered_cleaned_split.json --num-prompts 10 --request-rate 10
 ```
 
 ## Other baselines
@@ -42,14 +42,20 @@ python3 -m vllm.entrypoints.api_server --model meta-llama/Llama-2-7b-chat-hf --t
 
 ```
 # run synthetic
-python3 bench_throughput.py --backend vllm --port 30000 --tokenizer meta-llama/Llama-2-7b-chat-hf --num-prompt 1000 --request-rate 100 --input-len 1024 --output-len 256
+python3 bench_serving.py --backend vllm --port 30000 --tokenizer meta-llama/Llama-2-7b-chat-hf --num-prompt 1000 --request-rate 100 --input-len 1024 --output-len 256
 ```
 
 ```
 # run ShareGPT
-python3 bench_throughput.py --backend vllm --port 21000 --tokenizer meta-llama/Llama-2-7b-chat-hf --dataset ShareGPT_V3_unfiltered_cleaned_split.json --num-prompts 10 --request-rate 10
+python3 bench_serving.py --backend vllm --port 21000 --tokenizer meta-llama/Llama-2-7b-chat-hf --dataset ShareGPT_V3_unfiltered_cleaned_split.json --num-prompts 10 --request-rate 10
 ```
 
+```
+# run one batch
+python3 -m vllm.entrypoints.openai.api_server --model meta-llama/Meta-Llama-3-70B --tensor 8 --disable-log-requests --max-num-seqs 1024 --quantization fp8
+
+python3 bench_one.py --input-len 1024 --batch-size 1 1 2 4 8 16 32 64 128 256 512 768 1024 --port 8000 --backend vllm
+```
 
 ### LightLLM
 ```
@@ -57,5 +63,5 @@ python -m lightllm.server.api_server --model_dir ~/model_weights/Llama-2-7b-chat
 ```
 
 ```
-python3 bench_throughput.py --backend lightllm --port 22000 --tokenizer meta-llama/Llama-2-7b-chat-hf --dataset ShareGPT_V3_unfiltered_cleaned_split.json --num-prompts 10 --request-rate 10
+python3 bench_serving.py --backend lightllm --port 22000 --tokenizer meta-llama/Llama-2-7b-chat-hf --dataset ShareGPT_V3_unfiltered_cleaned_split.json --num-prompts 10 --request-rate 10
 ```
\ No newline at end of file

benchmark/latency_throughput/bench_one.py

@@ -15,19 +15,19 @@ def run_one_batch_size(bs):
     url = f"{args.host}:{args.port}"
     max_new_tokens = args.max_tokens
 
-    a = 20
-    prompt = f"{a, }"
+    if args.input_len:
+        input_ids = [
+            [int(x) for x in np.random.randint(0, high=16384, size=(args.input_len,))] for _ in range(bs)
+        ]
+    else:
+        text = [f"{i, }" for i in range(bs)]
 
     tic = time.time()
     if args.backend == "srt":
         if args.input_len:
-            inputs = {"input_ids": [
-                [int(x) for x in np.random.randint(0, high=16384, size=(args.input_len,))] for _ in range(bs)
-            ]}
+            inputs = {"input_ids": input_ids}
         else:
-            inputs = {"text": [
-                f"{i, }" for i in range(bs)
-            ]}
+            inputs = {"text": text}
 
         response = requests.post(
             url + "/generate",
@@ -44,7 +44,7 @@ def run_one_batch_size(bs):
         response = requests.post(
             url + "/generate",
             json={
-                "inputs": prompt,
+                "inputs": text[0],
                 "parameters": {
                     "temperature": 0,
                     "max_new_tokens": max_new_tokens,
@@ -53,13 +53,19 @@ def run_one_batch_size(bs):
             },
         )
     elif args.backend == "vllm":
+        if args.input_len:
+            inputs = {"prompt": input_ids}
+        else:
+            inputs = {"prompt": text}
+
         response = requests.post(
-            url + "/generate",
+            url + "/v1/completions",
             json={
-                "prompt": prompt,
+                "model": args.vllm_model_name,
                 "temperature": 0,
                 "max_tokens": max_new_tokens,
                 "ignore_eos": True,
+                **inputs,
             },
         )
     elif args.backend == "ginfer":
@@ -71,7 +77,7 @@ def run_one_batch_size(bs):
 
         tic = time.time()
         sample_request = sampler_pb2.SampleTextRequest(
-            prompt=prompt,
+            prompt=text[0],
             settings=sampler_pb2.SampleSettings(
                 max_len=max_new_tokens,
                 rng_seed=0,
@@ -92,7 +98,7 @@ def run_one_batch_size(bs):
     output_throughput = bs * max_new_tokens / latency
     print(f"latency: {latency:.2f} s, speed: {output_throughput:.2f} token/s")
 
-    with open("tmp_output.txt", "a") as fout:
+    with open("results.jsonl", "a") as fout:
         res = {
             "input_len": args.input_len,
             "output_len": args.max_tokens,
@@ -111,6 +117,7 @@ if __name__ == "__main__":
     parser.add_argument("--input-len", type=int, default=None)
     parser.add_argument("--batch-size", type=int, nargs='*', default=[1])
     parser.add_argument("--max-tokens", type=int, default=256)
+    parser.add_argument("--vllm-model-name", type=str, default="meta-llama/Meta-Llama-3-70B")
     args = parser.parse_args()
 
     if args.port is None:

python/sglang/README.md

+# Code Structure
+
+- `backend`: Various backends for the language interpreter.
+- `lang`: The frontend language.
+- `srt`: The runtime for running local models.
+- `test`: Test utilities.
+- `api.py`: Public API.
+- `bench_latency.py`: Benchmark utilities.
+- `global_config.py`: The global configs and constants.
+- `launch_server.py`: The entry point of launching local server.
+- `utils.py`: Common utilities.
+

python/sglang/srt/managers/controller/model_runner.py

@@ -276,17 +276,13 @@ class ModelRunner:
         input_metadata = InputMetadata.create(
             self,
             forward_mode=ForwardMode.EXTEND,
-            tp_size=self.tp_size,
             req_pool_indices=batch.req_pool_indices,
             seq_lens=batch.seq_lens,
             prefix_lens=batch.prefix_lens,
             position_ids_offsets=batch.position_ids_offsets,
             out_cache_loc=batch.out_cache_loc,
-            top_logprobs_nums=batch.top_logprobs_nums,
             return_logprob=batch.return_logprob,
-            flashinfer_prefill_wrapper_ragged=self.flashinfer_prefill_wrapper_ragged,
-            flashinfer_prefill_wrapper_paged=self.flashinfer_prefill_wrapper_paged,
-            flashinfer_decode_wrapper=self.flashinfer_decode_wrapper,
+            top_logprobs_nums=batch.top_logprobs_nums,
         )
         return self.model.forward(
             batch.input_ids,