Merge pull request #268 from AI-Hypercomputer:yuyan-prefix-cache-benchmark

PiperOrigin-RevId: 755894145

Author: jetstream authors

benchmarks/README.md

@@ -75,6 +75,7 @@ python benchmark_serving.py \
 ```
 
 ## Benchmark with openorca dataset (openorca is used by MLPerf inference for LLaMA2 models)
+
 ```
 python JetStream/benchmarks/benchmark_serving.py   \
 --tokenizer ~/maxtext/assets/tokenizer.llama2  \
@@ -93,6 +94,7 @@ python JetStream/benchmarks/benchmark_serving.py   \
 The benchmark has better performance if it first conducts a warmup of the JetStream server. We currently support `sampled` and `full` warmup modes. `full` mode would warmup up the JetStream server with all the input requests. `sampled` mode would warmup up the JetStream server with a sampling of the input requests across different bucket sizes of input lengths.
 
 Example to run benchmark with `full` warmup mode:
+
 ```
 python JetStream/benchmarks/benchmark_serving.py   \
 --tokenizer ~/maxtext/assets/tokenizer.llama2  \
@@ -115,7 +117,25 @@ python eval_accuracy.py outputs.json
 ```
 
 With openorca dataset and llama2-chat models (used by MLPerf), here are the reference accuracy numbers:
+
 ```
 llama2-7b-chat {'rouge1': 42.0706, 'rouge2': 19.8021, 'rougeL': 26.8474, 'rougeLsum': 39.5952, 'gen_len': 1146679, 'gen_num': 998}
 llama2-70b-chat {'rouge1': 44.4312, 'rouge2': 22.0352, 'rougeL': 28.6162}
-``` 
+```
+
+## Benchmark prefix cache
+
+Benchmark with mock input requests that share common prefix. Use to test prefix caching.
+
+All prompts length is `max-input-length`, and share common prefix mean at length `--prefix-cache-test-common-len` with normal distribution.
+
+```
+python JetStream/benchmarks/benchmark_serving.py \
+--tokenizer prefix_cache_test \
+--dataset prefix_cache_test
+--warmup-mode full \
+--num-prompts 100 \
+--max-input-length 16000 \
+--prefix-cache-test-common-len 9000\
+--max-output-length 50 \
+```

benchmarks/benchmark_prefix_cache.sh

@@ -0,0 +1,89 @@
+#!/bin/bash
+# Copyright 2025 Google LLC
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#      http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+set -e
+
+NUM_PROMPTS=${NUM_PROMPTS:-100}
+MAX_OUTPUT_LENGTH=${MAX_OUTPUT_LENGTH:-50}
+
+# Test combination from lengths and common prefix lengths.
+# The length should be shorter than max_input_length minus 1 for bos.
+BENCHMARK_PROMPT_LENGTHS=${BENCHMARK_PROMPT_LENGTHS:-8000,16000}
+BENCHMARK_PROMPT_COMMON_PREFIX_LENGTHS=${BENCHMARK_PROMPT_COMMON_PREFIX_LENGTHS:-4000,6000,8000,10000,12000,14000,16000}
+
+benchmark_serving_with_prefix_cache() {
+  echo "Starting prefix cache benchmark..."
+  echo "Benchmark serving script: ${BENCHMARK_SERVING_SCRIPT_PATH}"
+  echo "Prompt lengths to test: ${BENCHMARK_PROMPT_LENGTHS}"
+  echo "Common prefix lengths to test: ${BENCHMARK_PROMPT_COMMON_PREFIX_LENGTHS}"
+  echo "Number of prompts per run: ${NUM_PROMPTS}"
+  echo "Max output length per prompt: ${MAX_OUTPUT_LENGTH}"
+  echo "Base output directory for results: ${OUTPUTS_DIR_BASE}"
+  echo "Warmup mode: ${WARMUP_MODE}"
+
+  # Convert comma-separated strings to arrays for iteration
+  IFS=',' read -r -a prompt_lengths_arr <<< "$BENCHMARK_PROMPT_LENGTHS"
+  IFS=',' read -r -a common_prefix_lengths_arr <<< "$BENCHMARK_PROMPT_COMMON_PREFIX_LENGTHS"
+
+  for prompt_len in "${prompt_lengths_arr[@]}"; do
+    for common_len in "${common_prefix_lengths_arr[@]}"; do
+      if [ "${common_len}" -gt "${prompt_len}" ]; then
+        echo "Skipping: Common prefix length ${common_len} is greater than prompt length ${prompt_len}."
+        continue
+      fi
+
+      echo "----------------------------------------------------------------------"
+      echo "Running benchmark: Prompt Length=${prompt_len}, Common Prefix Length=${common_len}"
+      echo "----------------------------------------------------------------------"
+      echo "Warm up twice"
+      echo "----------------------------------------------------------------------"
+      
+      # With warmup-mode full, it will run twice
+      python3 ./benchmark_serving.py \
+        --tokenizer "prefix_cache_test" \
+        --dataset "prefix_cache_test" \
+        --num-prompts 10 \
+        --max-output-length "${MAX_OUTPUT_LENGTH}" \
+        --warmup-mode "full" \
+        --max-input-length "${prompt_len}" \
+        --prefix-cache-test-common-len "${common_len}"
+
+      echo "Warm up done"
+      echo "----------------------------------------------------------------------"
+
+      python3 ./benchmark_serving.py \
+        --tokenizer "prefix_cache_test" \
+        --dataset "prefix_cache_test" \
+        --num-prompts "${NUM_PROMPTS}" \
+        --max-output-length "${MAX_OUTPUT_LENGTH}" \
+        --warmup-mode "none" \
+        --max-input-length "${prompt_len}" \
+        --prefix-cache-test-common-len "${common_len}"
+
+      echo "Benchmark finished for Prompt Length=${prompt_len}, Common Prefix Length=${common_len}"
+      echo "----------------------------------------------------------------------"
+      echo
+    done
+  done
+  echo "All benchmark runs completed."
+}
+
+main() {
+  benchmark_serving_with_prefix_cache
+  echo "Script finished."
+  exit 0
+}
+
+main "$@"

benchmarks/benchmark_serving.py

@@ -209,6 +209,25 @@ def to_dict(self):
     }
 
 
+class PrefixCacheTestTokenizer:
+  """A simple tokenizer for testing prefix caching.
+
+  This tokenizer converts each character in a string to its integer ordinal
+  value during encoding, and converts a list of integer ordinals back to
+  a string during decoding. It's designed for testing scenarios, particularly
+  those involving prefix caching, where a basic, predictable tokenizer is
+  needed.
+  """
+
+  def encode(self, s: str, **kwargs) -> list[int]:
+    del kwargs
+    return [ord(c) for c in s]
+
+  def decode(self, token_ids: list[int], **kwargs) -> str:
+    del kwargs
+    return "".join([chr(token_id) for token_id in token_ids])
+
+
 def get_tokenizer(
     model_id: str,
     tokenizer_name: str,
@@ -219,6 +238,9 @@ def get_tokenizer(
   if tokenizer_name == "test":
     print("Using test tokenizer")
     return "test"
+  elif tokenizer_name == "prefix_cache_test":
+    print("Using prefix_cache_test tokenizer")
+    return PrefixCacheTestTokenizer()
   elif use_hf_tokenizer:
     # Please accept agreement to access private/gated models in HF, and
     # follow up instructions below to set up access token
@@ -329,6 +351,98 @@ def load_mmlu_dataset_csv(dataset_path: str) -> tuple[Any, dict[str, str]]:
   return combined_dataset, prompts_per_subject
 
 
+def load_mock_prefix_cache_test_input_requests(
+    prompt_len: int,
+    output_len: int,
+    common_prefix_len: int,
+    num_samples: int,
+) -> list[InputRequest]:
+  """Generates a mock dataset for testing prefix cache.
+
+  The prefix part of each prompt is a sub-string of a single master string.
+  The length of this prefix part for each sample is drawn from a normal
+  distribution with its mean set to `common_prefix_len`, and values are
+  clipped to the range [0, `prompt_len`].
+  The tokenizer is assumed to treat each character as a token.
+
+  Args:
+    prompt_len: The total length of each generated prompt string.
+    output_len: The length of each generated output string.
+    common_prefix_len: The target mean for the length of the prefix part
+                       of each prompt. These prefixes are derived from a
+                       shared master string.
+    num_samples: The number of (prompt, output) pairs to generate.
+
+  Returns:
+    A list of InputRequest objects.
+  """
+  if not 0 <= common_prefix_len <= prompt_len:
+    raise ValueError(
+        "Target mean common_prefix_len must be between 0 and prompt_len,"
+        f" inclusive. Got common_prefix_len={common_prefix_len}, "
+        f"prompt_len={prompt_len}"
+    )
+  if any(arg <= 0 for arg in [prompt_len, output_len, num_samples]):
+    raise ValueError(
+        "prompt_len, output_len, and num_samples cannot be 0 or negative."
+    )
+
+  input_requests: list[InputRequest] = []
+
+  # Generate a master string from which all prefixes will be derived.
+  # This ensures that prefixes of the same length are identical,
+  # and shorter prefixes are actual prefixes of longer ones.
+  master_potential_prefix = "".join(
+      random.choices("ABCDEFGHIJKLMNOPQRSTUVWXYZ", k=prompt_len)
+  )
+
+  # Generate prefix lengths for each sample from a normal distribution
+  scale = prompt_len / 3.0  # Standard deviation for the normal distribution
+
+  generated_prefix_lengths = np.random.normal(
+      loc=common_prefix_len, scale=scale, size=num_samples
+  )
+  generated_prefix_lengths = (
+      np.clip(generated_prefix_lengths, 0, prompt_len).round().astype(int)
+  )
+
+  for idx in range(num_samples):
+    current_actual_prefix_len = generated_prefix_lengths[idx]
+
+    actual_prefix_for_sample = master_potential_prefix[
+        :current_actual_prefix_len
+    ]
+
+    current_unique_len = prompt_len - current_actual_prefix_len
+    # This should not happen if generated_prefix_lengths is clipped correctly
+    if current_unique_len < 0:
+      current_unique_len = 0  # Safeguard
+      current_actual_prefix_len = prompt_len
+      actual_prefix_for_sample = master_potential_prefix[
+          :current_actual_prefix_len
+      ]
+
+    unique_suffix_str = "".join(
+        random.choices(
+            "abcdefghijklmnopqrstuvwxyz0123456789", k=current_unique_len
+        )
+    )
+
+    prompt_str = actual_prefix_for_sample + unique_suffix_str
+
+    output_str = "".join(random.choices("!@#$%^&*()_+", k=output_len))
+
+    request = InputRequest(
+        prompt=prompt_str,
+        prompt_len=len(prompt_str),
+        output=output_str,
+        output_len=len(output_str),
+        sample_idx=idx,
+    )
+    input_requests.append(request)
+  return input_requests
+
+
 def gen_mmlu_qa(data: Any, mmlu_method: str = "") -> str:
 
   output = ""
@@ -893,6 +1007,7 @@ def parse_args() -> argparse.Namespace:
           "mmlu",
           "math500",
           "longcontext",
+          "prefix_cache_test",
       ],
       help="The dataset name.",
   )
@@ -1086,6 +1201,12 @@ def parse_args() -> argparse.Namespace:
       choices=["HELM", "Harness", ""],
       help="mmlu method/format to generate shots",
   )
+  parser.add_argument(
+      "--prefix-cache-test-common-len",
+      type=int,
+      default=64,
+      help="Common prefix length for the prefix cache test dataset.",
+  )
   return parser.parse_args()
 
 
@@ -1112,6 +1233,13 @@ def main(args: argparse.Namespace):
     input_requests = mock_requests(
         args.total_mock_requests
     )  # e.g. [("AB", 2, "AB", 3)]
+  elif args.dataset == "prefix_cache_test":
+    input_requests = load_mock_prefix_cache_test_input_requests(
+        prompt_len=args.max_input_length,
+        output_len=args.max_output_length,
+        common_prefix_len=args.prefix_cache_test_common_len,
+        num_samples=args.num_prompts,
+    )
   else:
     dataset = []
     if args.dataset == "openorca":