nascheme

diff --git a/Include/internal/pycore_interp_structs.h b/Include/internal/pycore_interp_structs.h
index 6b3d5711b92..aadc577306f 100644
--- a/Include/internal/pycore_interp_structs.h
+++ b/Include/internal/pycore_interp_structs.h
@@ -229,7 +229,8 @@ struct _gc_runtime_state {
     PyObject *callbacks;

     Py_ssize_t heap_size;
-    Py_ssize_t work_to_do;
+    /* Total number of young objects since the last complete collection */

nascheme

import argparse
import gc
import sys
import time


def get_memory_usage():
    """Memory usage of the current process in KB."""
    result = {'peak': 0, 'rss': 0}
    with open('/proc/self/status') as status:

nascheme

Tested with https://github.com/nascheme/cpython/tree/gc-gen-3.14 (b4ea6d82324cf84f74731a88decd7577eccde2f9)
Benchmark script: https://gist.github.com/nascheme/6d09885d3696055b9665d3b88c7aacaa

inc_gc OFF (0) vs ON (1) comparison
   cycle    extra     live    t0(s)    t1(s)     Δt%     rss0     rss1   Δrss%    trash0    trash1
--------------------------------------------------------------------------------------------------
      10        0      100     0.66     0.73   +10.0      16M      17M      +6        6k       28k
      10        0     1.0k     0.66     0.76   +14.2      17M      20M     +19       14k       82k
      10        0    10.0k     0.78     0.84    +7.8      20M      26M     +29       88k      224k
      10     1.0k      100     0.73     0.76    +4.9      16M      20M     +21        6k       28k

nascheme

This is 3.14 branch with https://github.com/python/cpython/pull/142001 applied
Benchmark script: https://gist.github.com/nascheme/6d09885d3696055b9665d3b88c7aacaa


inc_gc OFF (0) vs ON (1) comparison
   cycle    extra     live    t0(s)    t1(s)     Δt%     rss0     rss1   Δrss%    trash0    trash1
--------------------------------------------------------------------------------------------------
      10        0      100     0.65     0.75   +16.3      16M      17M      +2        6k       11k
      10        0     1.0k     0.68     0.78   +14.3      17M      19M     +11       14k       54k
      10        0    10.0k     0.76     0.89   +17.8      20M      22M     +10       88k      138k

nascheme

base: d76c56e958c9a603ded42d27b39ab51c1e3794e4 (head of 3.14 branch)
gen_gc: https://github.com/nascheme/cpython/tree/gc-gen-3.14 (b4ea6d82324cf84f74731a88decd7577eccde2f9)
inc_gc: as above but with -Xinc_gc on

Benchmark results produced by "fastbench" 2ab489f1f315582b8a537b537a7664898d90a0a4


Benchmark                        base ms     gen_gc ms         ratio     inc_gc ms         ratio
------------------------------------------------------------------------------------------------
2to3                                95.2          96.6         1.015          97.7         1.026

nascheme

vLLM Tensor-Parallel Benchmark Profiler Analysis

Source: tp_generate_2026-10-25_2.json.gz — samply profiler output from TP (tensor-parallel) benchmark Model: LlamaForCausalLM (Llama architecture) vLLM version: v1 engine (inferred from vllm/v1/ paths) Executor: MultiprocExecutor (TP=2, 2 Worker_TP processes) Python: 3.14t (free-threaded build) Time window: 39s–100s (steady-state LLM inference) Total samples (in window): ~245,437

nascheme

vLLM Benchmark Profiler Analysis — CUDA Graphs Enabled

Source: mp_gen_cuda.json.gz — samply profiler output from mp_generate.py with --cuda-graphs enabled (multi-process, single-GPU-per-worker) Model: LlamaForCausalLM (Llama architecture, FP16 / c10::Half) vLLM version: v1 engine (inferred from vllm/v1/ paths) Time window: 56s–96s (steady-state LLM inference, excluding startup/shutdown) Total samples (in window): ~96,675

---

nascheme

# CUDA contention benchmark v2: realistic LLM-like workload.
#
# Simulates a transformer decode step more faithfully than v1:
# - Multiple linear layers (like attention Q/K/V projections + MLP)
# - Small element-wise ops (RMSNorm, activations, residual adds)
# - Index/gather operations (like token embedding lookup, KV cache scatter)
# - Tensor creation/destruction churn
# - Periodic CPU readback (like reading sampled token IDs)
#
# This should stress the same code paths as a real LLM step — many CUDA

nascheme

# Dual-engine multi-GPU threaded vLLM throughput benchmark.
#
# Architecture: two independent LLMEngine instances (one per GPU) fed from
# a shared tokenized-request queue, with a single tokenizer thread.
#
#   Tokenizer Thread (CPU)         Engine Thread 0 (cuda:0)   Engine Thread 1 (cuda:1)
#     input_processor.process()    add_request (from queue)   add_request (from queue)
#     tokenized_queue.put(ecr)     engine0.step()             engine1.step()
#                                  (continuous streaming)      (continuous streaming)
#

nascheme

# See: https://github.com/pytest-dev/pytest/issues/14077
from concurrent.futures import ThreadPoolExecutor
import numpy as np
import pytest

@pytest.fixture(scope="module")
def pool():
    yield ThreadPoolExecutor(max_workers=1)

def test_1(pool):