SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2606.00516 · ICML · 2026

Threshold-Based Exclusive Batching for LLM Inference

Yuzhou Nie, Bowen Pang, Weifang Zhang, Guangrui Ma, Shining Wu

arXiv · PDF · Open in the Atlas

Code that ran

7 of 13 samples (functions and classes) lifted from this paper's linked repositories executed on a synthesized input in Syntology's jail (0 of them constructed an object rather than computing a result; 0 honoured the contract we drafted, 0 violated it and 0 have no contract verdict; 7 failures of our instrument; 6 no recorded run), measured 2026-11-22; bodies inline only under a permissive licence as GitHub detects it. Not a reproduction of the paper's results.

"Ran" means the function executed on a synthesized input and returned a value. It is not a reproduction of the paper's results. This record, dated, as JSON: /api/ran/2606.00516.json · badge: /api/ran/2606.00516.svg

Each repository's role is a label: a classifier's output recorded on the link, or, where no role is recorded, what the Papers with Code archive marked. The line under it names its basis. Neither says who wrote the code.

RepositoryRoleRan
weifang231/eb-vllm canonical (classifier label)
classifier basis: census
7 ran · 6 no recorded run
FunctionStatusWhere it lives
bc_linter_include Ran
ran; our contract draft was wrong, not the code
weifang231/eb-vllm/vllm/_bc_linter.py
code served (permissive licence) · get_code("8549c45bd80a1ae5")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned function. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import torch.nn as nn

torch.manual_seed(0)

# The function bc_linter_include is a decorator that can be used with or without arguments.
# When called with obj=None (i.e., used as @bc_linter_include(reason="..."))
# it returns _wrap which wraps the decorated function.
# When called with an object directly, it just returns that object.

# Test case 1: Using as a decorator with reason parameter
def sample_function():
    """A sample public API function"""
    return "test"

# Test case 2: Using as a decorator without arguments
def another_function():
    """Another sample function"""
    pass

# Construct valid arguments
ARGS = {
    "obj": None,
    "reason": "This is a public API function that should be included in linting"
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type function — a function output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a function output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
bc_linter_skip Ran
ran; our contract draft was wrong, not the code
weifang231/eb-vllm/vllm/_bc_linter.py
code served (permissive licence) · get_code("e6eb11775bb350fd")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned function. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import torch.nn as nn
from typing import Any

torch.manual_seed(0)

# The function bc_linter_skip is a decorator that can be used with or without arguments.
# It takes:
# - obj: Any (optional, positional) - the object being decorated
# - reason: str | None (keyword-only, optional) - the reason for suppression

# Create a simple function to decorate
def legacy_api():
    """A legacy API function."""
    return "result"

# Create a simple class to decorate
class LegacyClass:
    """A legacy class."""
    pass

# The function can be called in multiple ways:
# 1. As @bc_linter_skip (no arguments)
# 2. As @bc_linter_skip() (called with no arguments)
# 3. As @bc_linter_skip(reason="...") (called with keyword argument)
# 4. With an object passed directly

ARGS = {
    "obj": legacy_api,
    "reason": "This is a legacy API that will be removed in future versions"
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type function — a function output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a function output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
get_beam_search_score Ran
ran; the contract check could not run (RAISES, DEP_MISSING or TIMEOUT)
weifang231/eb-vllm/vllm/beam_search.py
code served (permissive licence) · get_code("e2dd50c8a6072e9b")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned float, ndim 0, dtype family float, shape () as a later run recorded it. The image it ran in was not recorded. If your copy returns the same type, ndim and shape on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch

# Construct arguments for get_beam_search_score
tokens = [101, 2054, 2003, 102]  # Example token sequence with EOS token at end
cumulative_logprob = -5.2
eos_token_id = 102
length_penalty = 1.0

ARGS = {
    "tokens": tokens,
    "cumulative_logprob": cumulative_logprob,
    "eos_token_id": eos_token_id,
    "length_penalty": length_penalty,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type float, ndim 0, dtype family float, shape () — shape () as a later run recorded it
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
get_served_model_name Ran
ran; our contract draft was wrong, not the code
weifang231/eb-vllm/vllm/config/model.py
code served (permissive licence) · get_code("92fc6b49c51ae2c1")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned str. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import torch.nn as nn

torch.manual_seed(0)

# Test case 1: served_model_name is None (empty), should return model
model_1 = "gpt-3.5-turbo"
served_model_name_1 = None

# Test case 2: served_model_name is empty list, should return model
model_2 = "gpt-4"
served_model_name_2 = []

# Test case 3: served_model_name is non-empty string, should return served_model_name
model_3 = "claude"
served_model_name_3 = "custom-model-v1"

# Test case 4: served_model_name is non-empty list, should return first element
model_4 = "llama"
served_model_name_4 = ["served-v1", "served-v2", "served-v3"]

# Use test case 3 (non-empty string) as the primary test
ARGS = {
    "model": "gpt-3.5-turbo",
    "served_model_name": "custom-served-model"
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type str — a str output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a str output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
run Ran
ran; our contract draft was wrong, not the code
weifang231/eb-vllm/vllm/collect_env.py
code served (permissive licence) · get_code("bd5675c0d72d5295")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned tuple, ndim 1, dtype family float; no output shape is stated. The image it ran in was not recorded. If your copy returns the same type and ndim on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import locale
import subprocess
import sys

def get_platform():
    if sys.platform.startswith('linux'):
        return 'linux'
    elif sys.platform.startswith('win32'):
        return 'win32'
    elif sys.platform.startswith('cygwin'):
        return 'cygwin'
    elif sys.platform.startswith('darwin'):
        return 'darwin'
    else:
        return sys.platform

def run(command):
    """Return (return-code, stdout, stderr)."""
    shell = True if type(command) is str else False
    try:
        p = subprocess.Popen(command, stdout=subprocess.PIPE, stderr=subprocess.PIPE, shell=shell)
        raw_output, raw_err = p.communicate()
        rc = p.returncode
        if get_platform() == 'win32':
            enc = 'oem'
        else:
            enc = locale.getpreferredencoding()
        output = raw_output.decode(enc)
        if command == 'nvidia-smi topo -m':
            output = output.rstrip()
        else:
            output = output.strip()
        err = raw_err.decode(enc)
        return (rc, output, err.strip())
    except FileNotFoundError:
        cmd_str = command if isinstance(command, str) else command[0]
        return (127, '', f'Command not found: {cmd_str}')


# Test with a simple string command
ARGS = {
    'command': 'echo hello'
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
none outside the standard library — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type tuple, ndim 1, dtype family float — a tuple output has no array shape of its own; none is stated
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
run_and_parse_first_match Ran
ran; our contract draft was wrong, not the code
weifang231/eb-vllm/vllm/collect_env.py
code served (permissive licence) · get_code("0b342669284f2be6")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned str. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import regex as re

def mock_run_lambda(command):
    """Mock run_lambda that returns (return_code, output, error)"""
    # Simulate successful command execution with output containing a match
    if command == "test_command":
        return (0, "version: 1.2.3", None)
    else:
        return (1, "", None)

ARGS = {
    "run_lambda": mock_run_lambda,
    "command": "test_command",
    "regex": r"version: ([\d.]+)"
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
regex — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type str — a str output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a str output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
run_and_read_all Ran
ran; our contract draft was wrong, not the code
weifang231/eb-vllm/vllm/collect_env.py
code served (permissive licence) · get_code("5ae630bfc3821b6e")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned str. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import torch.nn as nn

def create_run_lambda():
    """Create a mock run_lambda function that returns (rc, stdout, stderr)"""
    def mock_run_lambda(command):
        # Simulate successful command execution
        if "error" in command:
            return 1, "", "error occurred"
        else:
            return 0, "command output", ""
    return mock_run_lambda

ARGS = {
    "run_lambda": create_run_lambda(),
    "command": "echo 'test command'"
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type str — a str output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a str output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
convert_vertical_slash_indexes Not yet run
unverified (harvested, no recorded run)
weifang231/eb-vllm/vllm/_custom_ops.py
code served (permissive licence) · get_code("89f4968f0bf052d7")
How it ran

no invocation recorded.

convert_vertical_slash_indexes_mergehead Not yet run
unverified (harvested, no recorded run)
weifang231/eb-vllm/vllm/_custom_ops.py
code served (permissive licence) · get_code("8c9d2bd2211f1bda")
How it ran

no invocation recorded.

get_graph_partition_signature_patched Not yet run
unverified (harvested, no recorded run)
weifang231/eb-vllm/vllm/env_override.py
code served (permissive licence) · get_code("41329a8dffd43d3a")
How it ran

no invocation recorded.

memory_plan_reuse_patched Not yet run
unverified (harvested, no recorded run)
weifang231/eb-vllm/vllm/env_override.py
code served (permissive licence) · get_code("2842ff1da04c998d")
How it ran

no invocation recorded.

rms_norm_dynamic_per_token_quant Not yet run
unverified (harvested, no recorded run)
weifang231/eb-vllm/vllm/_custom_ops.py
code served (permissive licence) · get_code("8954fe77978adce8")
How it ran

no invocation recorded.

should_partition_patched Not yet run
unverified (harvested, no recorded run)
weifang231/eb-vllm/vllm/env_override.py
code served (permissive licence) · get_code("7793c3aaf91b92fd")
How it ran

no invocation recorded.

Also present in other papers' code

Of 13 samples lifted from this paper's linked repositories, 13 carry a function identity and 0 do not. Of the 13 distinct functions with an identity, 9 are also present in other papers whose code we hold (7 with at least one paper dated earlier, 0 the same month, 7 later) and 4 are only here among the 52,977 papers whose code Syntology holds (as of 2026-09-26).

Only here means: among the 52,977 papers whose code Syntology holds (51,887 with lifted samples, 1,090 through their assemblies only), as of 2026-09-26, no other paper's lifted samples or current assembly bricks contain an identical function body after normalisation.

FunctionAlso present inEarliest-datedOthers named
run_and_parse_first_match 14 other papers
13 dated earlier, 1 later
2109.04003 (2021-09)
Improving Deep Metric Learning by Divide and Conquer
2109.08267 (2021-09)
CompilerGym: Robust, Performant Compiler Optimization Environments for AI Research
2202.04053 (2022-02)
DALL-Eval: Probing the Reasoning Skills and Social Biases of Text-to-Image Generation Models
and 11 more not named here
run_and_read_all 14 other papers
13 dated earlier, 1 later
2109.04003 (2021-09)
Improving Deep Metric Learning by Divide and Conquer
2109.08267 (2021-09)
CompilerGym: Robust, Performant Compiler Optimization Environments for AI Research
2202.04053 (2022-02)
DALL-Eval: Probing the Reasoning Skills and Social Biases of Text-to-Image Generation Models
and 11 more not named here
get_graph_partition_signature_patched 2 other papers
1 dated earlier, 1 later
2309.06180 (2023-09)
Efficient Memory Management for Large Language Model Serving with PagedAttention
2608.19920 (2026-08)
Learning how to Forget: Fine-tuning for Long-Context Sparse Attention
get_served_model_name 2 other papers
1 dated earlier, 1 later
2309.06180 (2023-09)
Efficient Memory Management for Large Language Model Serving with PagedAttention
2608.19920 (2026-08)
Learning how to Forget: Fine-tuning for Long-Context Sparse Attention
memory_plan_reuse_patched 2 other papers
1 dated earlier, 1 later
2309.06180 (2023-09)
Efficient Memory Management for Large Language Model Serving with PagedAttention
2608.19920 (2026-08)
Learning how to Forget: Fine-tuning for Long-Context Sparse Attention
run 2 other papers
1 dated earlier, 1 later
2309.06180 (2023-09)
Efficient Memory Management for Large Language Model Serving with PagedAttention
2608.19920 (2026-08)
Learning how to Forget: Fine-tuning for Long-Context Sparse Attention
should_partition_patched 2 other papers
1 dated earlier, 1 later
2309.06180 (2023-09)
Efficient Memory Management for Large Language Model Serving with PagedAttention
2608.19920 (2026-08)
Learning how to Forget: Fine-tuning for Long-Context Sparse Attention
convert_vertical_slash_indexes 1 other paper
1 undated
no dated paper among those named
openreview_ky9RAeT6ie
MOSAIC: Unlocking Over 30× Context Length for Diffusion LLMs Inference via Global Memory Planning and Dynamic Peak Taming
convert_vertical_slash_indexes_mergehead 1 other paper
1 undated
no dated paper among those named
openreview_ky9RAeT6ie
MOSAIC: Unlocking Over 30× Context Length for Diffusion LLMs Inference via Global Memory Planning and Dynamic Peak Taming

Presence, not direction: an identical function body in two papers' repositories does not say who wrote it, which came from which, or that either paper's method rests on it. Matching is exact after normalisation, so a renamed or edited function is not matched. Dates are when papers appeared, not when code was written. Only .py files the harvester lifted are covered; notebooks are not read. A new-style arXiv id YYMM.NNNN dates a paper to the month 20YY-MM; otherwise Paper.year dates it to the year; otherwise it is undated. Dates are when the paper appeared, not when the code was written. Read 2026-11-22.

Repositories linked to this paper

Some links come from the archived Papers with Code dataset (CC BY-SA 4.0): attribution and licence.

Abstract

Mixed batching (MB)-interleaving prefill and decode in a single batch-has become the standard scheduling strategy for large language model (LLM) inference due to its efficiency in maximizing compute and memory utilization. However, through controlled experiments, we find that prefill-decode interference inflates MB's per-step marginal cost above that of pure decode. On the high-bandwidth H200 (4.8 TB/s), this occurs only when decode tokens exceed 80% of the batch; however, on the bandwidth-constrained RTX PRO 6000 (1.792 TB/s), this threshold plummets to just 20%. Consequently, the optimal choice between MB and exclusive batching (EB) fundamentally depends on GPU memory bandwidth, model size, and workload composition. We derive a closed-form condition for this EB-MB performance crossover, along with asymptotically optimal phase-switching thresholds and memorysafe batch sizing for EB. Optimized EB achieves up to 41.9% higher throughput on bandwidthconstrained GPUs, while MB retains its advantage on high-bandwidth hardware with larger models. Our hybrid scheduler EB + applies this condition online to dynamically switch between EB and MB without manual intervention. Under nonstationary traffic with distribution or concurrency shifts, EB + attains the highest or near-highest throughput in every setting, outperforming MB by up to 36.4%. 1

For agents

The same record, over MCP at https://syntology.ai/mcp:

get_harvested_code_for_paper("2606.00516")
get_code_for_paper("2606.00516")
have("2606.00516")

The run record, dated, one paper per request, free:

curl https://syntology.ai/api/ran/2606.00516.json

A badge for a README (the split and the date, never a ratio):

[![Syntology run record](https://syntology.ai/api/ran/2606.00516.svg)](https://syntology.ai/paper/2606.00516)

Connect an agent — have() is free.