Skip to content

Support for spin in attention benchmark and better handling of max length #1001

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
copybara-service merged 1 commit into from
Aug 31, 2026
+16 −55
Merged

Support for spin in attention benchmark and better handling of max length #1001

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
71 changes: 16 additions & 55 deletions evals/attention_benchmark.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,4 +1,4 @@
// Copyright 2026 Google LLC
// Copyright 2024 Google LLC
// SPDX-License-Identifier: Apache-2.0
//
// Licensed under the Apache License, Version 2.0 (the "License");
Expand All @@ -13,46 +13,11 @@
// See the License for the specific language governing permissions and
// limitations under the License.

// Standalone benchmark tool for Gemma attention implementations.
//
// This binary benchmarks different attention implementations (e.g., standard
// flash attention, tiled attention, scalar attention) across various sequence
// lengths and batch sizes by simulating history in the KV cache and measuring
// decode generation speed.
//
// =============================================================================
// Usage Example
// =============================================================================
//
// bazel run -c opt //third_party/gemma_cpp:attention_benchmark -- \
// --weights /path/to/weights.sbs \
// --tokenizer /path/to/tokenizer.spm \
// --attention_impl tiled_flash \
// --context_lengths 1024,4096,8192 \
// --prefill_length 1 \
// --benchmark_tokens 32 \
// --benchmark_batch_size 1
//
// =============================================================================
// Supported Benchmark Flags
// =============================================================================
// --benchmark_mode: Benchmark mode: 'both' (benchmarks prefill and
// decode) or 'decode' (benchmarks decode with simulated
// KV cache history). Default: 'both'.
// --context_lengths: Comma-separated KV cache history lengths to
// benchmark (default: '1024,8192,32768'; aliases:
// --generation_context_lengths, --prompt_lengths)
// --prefill_length: Number of prompt tokens before decode generation
// (default: 1; alias: --generation_prefill_length)
// --benchmark_tokens: Number of decode tokens to generate (default: 32)
// --benchmark_batch_size: Number of queries to benchmark in a batch
// (default: 1)

#include <stddef.h>

#include <algorithm>
// Standalone benchmark measuring single-token autoregressive decoding
// generation throughput across simulated KV cache history lengths.

#include <cstddef>
#include <cstdlib>
#include <cstring>
#include <iostream>
#include <string>
#include <vector>
Expand All @@ -76,6 +41,8 @@ class AttentionBenchmarkArgs : public ArgsBase<AttentionBenchmarkArgs> {
InitAndParse(argc, argv, consumed);
}

AttentionBenchmarkArgs() = default;

std::string context_lengths;
std::string generation_context_lengths;
std::string prompt_lengths;
Expand Down Expand Up @@ -205,6 +172,7 @@ int main(int argc, char** argv) {
gcpp::RuntimeConfig runtime_config{};
args.inference.CopyTo(runtime_config);
runtime_config.verbosity = args.inference.verbosity;
runtime_config.use_spinning = args.threading.spin;
size_t decode_tokens = bench_args.benchmark_tokens;
size_t num_queries = bench_args.benchmark_batch_size;
if (num_queries == 0) {
Expand Down Expand Up @@ -251,17 +219,12 @@ int main(int argc, char** argv) {
continue;
}

const size_t max_seq_len = gemma.Config().max_seq_len;
const size_t priming_tokens = is_only_decode ? prefill_len : 0;
if (target_len + priming_tokens + decode_tokens > max_seq_len) {
const size_t max_allowed =
(max_seq_len > priming_tokens + decode_tokens)
? max_seq_len - priming_tokens - decode_tokens
: 0;
std::cerr << "Warning: target_len=" << target_len
<< " exceeds max_seq_len (" << max_seq_len << "), clamping to "
<< max_allowed << "\n";
target_len = max_allowed;
const size_t total_capacity =
target_len + priming_tokens + decode_tokens + 32;
if (total_capacity > gemma.Config().max_seq_len) {
const_cast<gcpp::ModelConfig&>(gemma.Config())
.SetMaxSeqLen(total_capacity);
}

const size_t history_len = is_only_decode ? target_len : 0;
Expand All @@ -272,9 +235,6 @@ int main(int argc, char** argv) {
: "Context / Prefill Length: ")
<< target_len << " tokens\n";

size_t total_capacity =
std::min(max_seq_len, history_len + prompt_tokens + decode_tokens + 32);

size_t original_seq_len = args.inference.seq_len;
args.inference.seq_len = total_capacity;

Expand All @@ -283,7 +243,8 @@ int main(int argc, char** argv) {
gen_config.decode_qbatch_size = num_queries;
gen_config.sample_func = [](size_t, size_t, gcpp::Logits,
size_t) -> gcpp::TokenAndProb {
// Return an arbitrary non-EOS token ID so benchmarks never terminate early.
// Return an arbitrary non-EOS token ID so benchmarks never
// terminate early.
return gcpp::TokenAndProb{500, 1.0f};
};

Expand All @@ -299,7 +260,7 @@ int main(int argc, char** argv) {

std::vector<int> tokens = GenerateSyntheticPrompt(gemma, prompt_tokens);

gcpp::TimingInfo timing_info;
gcpp::TimingInfo timing_info{.verbosity = args.inference.verbosity};

size_t generated = 0;
const size_t total_prefill_tokens = num_queries * prompt_tokens;
Expand Down
Loading