which have pervaded practically each aspect of our day by day lives are autoregressive decoder fashions. These fashions apply compute-heavy kernel operations to churn out tokens one after the other in a way that, at first look, appears extraordinarily inefficient. Given the large demand for generative AI, it’s no shock that extraordinary engineering effort is being invested into its optimization. Whether or not it’s by means of customized CUDA kernels, CUDA Graphs, devoted AI accelerators, or speculative sampling — any method that reduces latency and/or price by even a fraction of a share is a win.
On this publish, we exhibit a method for optimizing token technology in PyTorch utilizing CUDA stream interleaving. Whereas easy to implement, the tactic addresses a selected, typically neglected bottleneck and might result in significant efficiency boosts. Whereas pipelining mannequin execution utilizing CUDA streams is frequent in AI techniques engineering, we didn’t discover any tutorial documenting the precise PyTorch-level software we describe right here. For those who discover the method helpful, please be so sort as to reference this publish.
To facilitate our dialogue, we’ll use a easy GPT-2 PyTorch decoder mannequin from HuggingFace’s transformers (v5.1.0) library. We’ll run our experiments on an NVIDIA L40S GPU and PyTorch (2.10.0).
Disclaimer: The code we’ll share is meant for demonstrative functions. Please don’t depend on its accuracy or optimality. Please don’t interpret our mentions of any library, platform, or service as an endorsement of its use.
Importantly, the worth of the CUDA stream-based technique we’ll focus on can range vastly primarily based on the small print of your mannequin and runtime setting. Please be sure you run your personal benchmarks earlier than integrating its use.
Our focus on this publish is on PyTorch-native inference workloads which stay extraordinarily prevalent in growth and take a look at settings. Nonetheless, it is very important observe that for manufacturing environments devoted LLM inference libraries akin to vLLM or NVIDIA TensorRT-LLM are likely to ship higher efficiency and ought to be used every time related.
A Toy GPT-2 Mannequin
To simplify our dialogue, we’ll use a GPT-2 decoder mannequin from the HuggingFace transformers library and have it run autoregressively on a batch of empty prompts.
Within the following code block, we initialize the mannequin and outline a naive token technology operate that creates a batch of random streams as much as a given size.
import torch
from transformers import GPT2LMHeadModel, GPT2Config
torch.set_float32_matmul_precision('excessive')
DEVICE = "cuda"
# outline the decoder mannequin
config = GPT2Config.from_pretrained("gpt2")
mannequin = GPT2LMHeadModel(config).to(DEVICE).eval()
@torch.inference_mode()
def generate_sequence(mannequin, max_seqlen, batch_size):
# Initialize prompts with BOS token
all_tokens = torch.full(
(batch_size, 1),
config.bos_token_id,
machine=DEVICE,
dtype=torch.lengthy
)
completed = torch.zeros(batch_size, machine=DEVICE, dtype=torch.bool)
for i in vary(max_seqlen):
outputs = mannequin(all_tokens)
# extract new token
logits = outputs.logits[:, -1, :]
new_tokens = torch.argmax(logits, dim=-1)
# append new token to sequence
all_tokens = torch.cat(
[all_tokens, new_tokens.unsqueeze(-1)],
dim=-1
)
completed |= (new_tokens == config.eos_token_id)
stop_gpu = torch.all(completed)
# checking cease situation
if stop_gpu.merchandise():
print(f"All sequences completed at step {i+1}")
break
return all_tokensSubsequent, we outline a easy benchmarking operate which we use to measure the runtime efficiency and reminiscence utilization of our token generator in several eventualities.
import time, statistics
def benchmark(func, num_runs=10):
# Warmup
func()
torch.cuda.synchronize()
runtimes = []
for _ in vary(num_runs):
# reset reminiscence stats earlier than every run
torch.cuda.empty_cache()
torch.cuda.reset_peak_memory_stats()
torch.cuda.synchronize()
begin = time.perf_counter()
_ = func()
torch.cuda.synchronize()
finish = time.perf_counter()
runtimes.append(finish - begin)
# Get reminiscence allocator stats from final run
mem_stats = torch.cuda.memory_stats()
allocated_peak = mem_stats.get('allocated_bytes.all.peak', 0)
reserved_peak = mem_stats.get('reserved_bytes.all.peak', 0)
f_peak = reserved_peak - allocated_peak
f_pct = (
100 * f_peak / reserved_peak
if reserved_peak > 0 else 0
)
print(f"n{'='*60}")
print(f"Runtime Outcomes:")
print(f" Imply: {statistics.imply(runtimes):.4f}s")
print(f" Std: {statistics.stdev(runtimes):.4f}s")
print(f" Min: {min(runtimes):.4f}s")
print(f" Max: {max(runtimes):.4f}s")
print(f"nMemory Stats:")
print(f" Allotted bytes (peak): {allocated_peak / 1e9:.3f} GB")
print(f" Reserved bytes (peak): {reserved_peak / 1e9:.3f} GB")
print(f" Fragmentation (peak): {f_peak / 1e9:.3f} GB ({f_pct:.1f}%)")
print(f"{'='*60}n")
batch_size = 32
for max_seqlen in [100, 200, 400]:
print(
f"Benchmarking technology with batch dimension {batch_size} "
f"and max sequence size {max_seqlen}..."
)
benchmark(
lambda: generate_sequence(
mannequin, max_seqlen=max_seqlen, batch_size=batch_size
)
)Within the desk beneath we seize the outcomes for a batch dimension of 32 and several other completely different sequence lengths:
Because the sequence size doubles, the runtime quadruples — showing to observe a traditional O(N²) scaling sample. Moreover, excessive reminiscence fragmentation factors to extreme pressure on the CUDA reminiscence allocator, which may end up in frequent reminiscence faults and degrade runtime efficiency. The fragmentation outcomes from every step asking for barely bigger tensor allocations, a sample which finally ends up leaving a number of pockets of unusable reminiscence.
Our first optimization, KV caching, addresses the runtime complexity of our decoder mannequin.
KV Caching
Our naive generator is extraordinarily inefficient — quite than storing and reusing the intermediate tensors from earlier tokens, it recalculates your entire sequence at each step.
We handle the computation inefficiency by utilizing KV caching: We retailer and reuse the intermediate Key and Worth tensors for earlier tokens. KV caching reduces the runtime complexity of token technology from O(N²) to O(N).
Within the following code block, we make the most of the transformers library’s built-in assist for KV caching to reprogram our token technology operate to compute a single batch of tokens in every step.
@torch.inference_mode()
def generate_sequence(mannequin, max_seqlen, batch_size, use_cache=False):
# Initialize prompts with BOS token
all_tokens = torch.full(
(batch_size, 1),
config.bos_token_id,
machine=DEVICE,
dtype=torch.lengthy
)
completed = torch.zeros(batch_size, machine=DEVICE, dtype=torch.bool)
# past_key_values is used to retailer the cached key/values for every layer
past_key_values = None
for i in vary(max_seqlen):
current_input = (
all_tokens if past_key_values is None
else all_tokens[:, -1:]
)
outputs = mannequin(
current_input,
past_key_values=past_key_values,
use_cache=use_cache
)
# replace cache for subsequent step
past_key_values = outputs.past_key_values
logits = outputs.logits[:, -1, :]
new_tokens = torch.argmax(logits, dim=-1)
# append new token to sequence
all_tokens = torch.cat(
[all_tokens, new_tokens.unsqueeze(-1)],
dim=-1
)
completed |= (new_tokens == config.eos_token_id)
stop_gpu = torch.all(completed)
# checking cease situation
if stop_gpu.merchandise():
print(f"All sequences completed at step {i+1}")
break
return all_tokensThe ensuing efficiency numbers are captured within the following desk:
The efficiency enchancment is profound and, as anticipated, will increase as a operate of the sequence size.
Though considerably higher than in our baseline experiment, the diploma of reminiscence fragmentation stays a priority. To handle this we discover two strategies, expandable reminiscence allocations and static KV caching.
Expandable CUDA Reminiscence Allocations
To scale back CUDA reminiscence fragmentation, we program PyTorch to make use of expandable memory segments. As of the time of this writing, this reminiscence optimization is an experimental characteristic and ought to be used with warning. Please see the PyTorch documentation for particulars. To make use of the characteristic we set the next setting variable:
export PYTORCH_ALLOC_CONF="expandable_segments:True"Rerunning our benchmark leads to the next desk:
Not solely can we see a marked enchancment in fragmentation, however we additionally get an extra (marginal) enchancment in runtime efficiency.
KV Caching With StaticCache
The default cache in HuggingFace is dynamic — it grows because the variety of keys and values will increase through the technology progresses. HuggingFace helps a fixed-size cache, StaticCache, which pre-allocates a most cache dimension for the KV pairs and reduces pressure on the CUDA reminiscence allocator. The drawback of utilizing StaticCache is that the complete size of the cache participates within the consideration computation at every token technology step, the place irrelevant tokens are masked out. This leads to a waste of computation that grows with the sequence size. For instance, when producing a sequence of 400 tokens, the eye computation for every token might be run on full 400X400-sized tensors.
Within the code block beneath we improve our sequence generator to assist the usage of a StaticCache:
che:
from transformers import StaticCache
@torch.inference_mode()
def generate_sequence(
mannequin, max_seqlen, batch_size, use_cache=False, use_static_cache=False
):
# Initialize prompts with BOS token
all_tokens = torch.full(
(batch_size, 1),
config.bos_token_id,
machine=DEVICE,
dtype=torch.lengthy
)
completed = torch.zeros(batch_size, machine=DEVICE, dtype=torch.bool)
# Initialize static cache if requested
if use_cache and use_static_cache:
past_key_values = StaticCache(
config=config,
max_batch_size=batch_size,
max_cache_len=max_seqlen,
machine=DEVICE,
dtype=mannequin.dtype
)
else:
past_key_values = None
# Initialize cache place monitoring for static cache
cache_positions = torch.arange(max_seqlen, machine=DEVICE)
for i in vary(max_seqlen):
current_input = (
all_tokens if past_key_values is None
else all_tokens[:, -1:]
)
cache_position = (
cache_positions[i:i+1] if use_static_cache else None
)
outputs = mannequin(
current_input,
past_key_values=past_key_values,
cache_position=cache_position,
use_cache=use_cache
)
# replace cache for subsequent step
past_key_values = outputs.past_key_values
logits = outputs.logits[:, -1, :]
new_tokens = torch.argmax(logits, dim=-1)
# append new token to sequence
all_tokens = torch.cat(
[all_tokens, new_tokens.unsqueeze(-1)],
dim=-1
)
completed |= (new_tokens == config.eos_token_id)
stop_gpu = torch.all(completed)
# checking cease situation
if stop_gpu.merchandise():
print(f"All sequences completed at step {i+1}")
break
return all_tokensThe up to date outcomes are captured beneath:
Utilizing a fixed-sized cache vastly improves reminiscence utilization as indicated by the lower in reminiscence fragmentation. Nonetheless, its influence on runtime efficiency is combined — for 100 tokens it reduces efficiency in comparison with a dynamic cache, whereas for 200 and 400 tokens it boosts efficiency by 9% and 10%, respectively.
There are extra superior strategies of implementing consideration that optimize for reminiscence utilization with out the price of wasted computation. In a earlier publish, Optimizing Transformer Fashions for Variable-Size Enter Sequences, we lined some PyTorch methods for computing consideration sparsely to scale back computation waste. For manufacturing settings, libraries akin to vLLM use PagedAttention for maximizing reminiscence utilization. These strategies are outdoors the scope of this publish.
For extra particulars on caching in HuggingFace, please see the caching strategies overview.
Mannequin Compilation
One of many documented benefits of utilizing a fixed-sized cache is that it permits for making the most of many just-in-time (JIT) optimizations.
Within the following code block we apply our benchmark to a PyTorch-compiled model of our decoder mannequin:
batch_size = 32
max_seqlen = 100
mannequin = torch.compile(mannequin)
benchmark(
lambda: generate_sequence(
mannequin,
max_seqlen=max_seqlen,
batch_size=batch_size,
use_cache=True,
use_static_cache=True
)
)Mannequin compilation leads to an extra increase to runtime efficiency as proven within the desk beneath:
Observe that we are able to apply mannequin compilation when utilizing dynamic caching, as effectively. Nonetheless, torch.compile supplies the very best outcomes when the computation graph consists of fixed-sized tensors (e.g., see right here for extra particulars).
The Efficiency Penalty of Early Stopping
An integral a part of frequent token turbines is checking for the end-of-sequence (EOS) on the finish of every step. With out this take a look at, token turbines would at all times run for max_seqlen, even when all of the sequences within the batch have ended. This might lead to appreciable computation waste and pointless latency — particularly when frequent sequence lengths are a lot shorter than the utmost size. Within the case of our toy experiment, we look forward to all of the sequences within the batch to finish and discontinue token technology. Manufacturing-grade implementations will generally carry out steady batching — changing accomplished sequences with new prompts on the enter queue.
completed |= (new_tokens == config.eos_token_id)
stop_gpu = torch.all(completed)
# checking cease situation
if stop_gpu.merchandise():
print(f"All sequences completed at step {i+1}")
breakImportantly, the .merchandise() name on the stop_gpu tensor, triggers a blocking host-device synchronization occasion. Extra particularly, so as to consider the conditional if assertion, the CPU should look forward to the GPU to finish its computation and replica the contents of the tensor to host reminiscence. Whereas the CPU waits, it’s blocked from executing the following step of the token technology loop, or extra precisely, it’s blocked from loading the following computation kernels onto the GPU.
To measure the influence of the stopping situation on runtime efficiency, we add instrumentation for efficiency profiling with NVIDIA Nsight™ Systems (nsys) utilizing the torch.cuda.profiler and nvtx (v0.2.14) APIs. (See our current publish for extra particulars on efficiency profiling with nsys).
ore particulars on efficiency profiling with nsys).
import nvtx
from torch.cuda import profiler
@torch.inference_mode()
def generate_sequence(
mannequin, max_seqlen, batch_size, use_cache=False, use_static_cache=False
):
# Initialize prompts with BOS token
all_tokens = torch.full(
(batch_size, 1),
config.bos_token_id,
machine=DEVICE,
dtype=torch.lengthy
)
completed = torch.zeros(batch_size, machine=DEVICE, dtype=torch.bool)
# Initialize static cache if requested
if use_cache and use_static_cache:
past_key_values = StaticCache(
config=config,
max_batch_size=batch_size,
max_cache_len=max_seqlen,
machine=DEVICE,
dtype=mannequin.dtype
)
else:
past_key_values = None
# Initialize cache place monitoring for static cache
cache_positions = torch.arange(max_seqlen, machine=DEVICE)
for i in vary(max_seqlen):
if i == 30:
# begin nsys profiler
torch.cuda.synchronize()
profiler.begin()
elif i == 50:
# cease nsys profiler
torch.cuda.synchronize()
profiler.cease()
with nvtx.annotate(f"Step {i+1}", coloration="blue"):
with nvtx.annotate("Mannequin Ahead", coloration="inexperienced"):
current_input = (
all_tokens if past_key_values is None
else all_tokens[:, -1:]
)
cache_position = (
cache_positions[i:i+1] if use_static_cache else None
)
outputs = mannequin(
current_input,
past_key_values=past_key_values,
cache_position=cache_position,
use_cache=use_cache
)
past_key_values = outputs.past_key_values
logits = outputs.logits[:, -1, :]
new_tokens = torch.argmax(logits, dim=-1)
all_tokens = torch.cat(
[all_tokens, new_tokens.unsqueeze(-1)],
dim=-1
)
completed |= (new_tokens == config.eos_token_id)
stop_gpu = torch.all(completed)
with nvtx.annotate("Test Cease Situation", coloration="crimson"):
# checking cease situation
if stop_gpu.merchandise():
print(f"All sequences completed at step {i+1}")
break
return all_tokensWe run our script utilizing the cudaProfilerApi possibility to start out and cease the profiler programmatically. Please see the official documentation for full particulars on profiling from the nsys CLI.
nsys profile
--capture-range=cudaProfilerApi
--trace=cuda,nvtx,osrt
--output=baseline
python practice.pyThe next hint, captured for a batch dimension of 16 and sequence size of 100, reveals the GPU idling for about 110 microseconds in between steps — an eternity within the context of high-performance GPU workloads. It is a direct results of the synchronization occasion triggered by the EOS take a look at.
In production-grade implementations such synchronization points are prevented by some mixture of 1) use of decrease degree (e.g., C/C++) code that avoids the limitation of the Python interpreter, 2) utilizing CUDA graphs to scale back overhead of kernel loading, 3) shifting conditional checks onto the GPU utilizing conditional nodes, and 4) repeatedly and asynchronously making ready subsequent requests whereas the EOS examine is in progress.
Within the subsequent part, we exhibit a method for hiding the overhead of the host-device synchronization in PyTorch utilizing CUDA streams.
A CUDA Stream Optimization
A CUDA stream is a linear sequence of operations (kernels, reminiscence copies, and many others.) that execute so as on the GPU. Whereas operations inside a single stream are assured to execute sequentially, operations in several streams can execute concurrently or overlap.
In earlier posts (e.g., right here and right here) we demonstrated the usage of CUDA streams in pipelining frequent AI/ML workloads, e.g., executing a mannequin on batch N whereas making ready batch N+1. On this publish we’ll use CUDA streams to allow the CPU to load the GPU kernels of step N+1 earlier than checking the stopping standards of step N. Opposite to our earlier demonstrations of CUDA streams, our present instance won’t essentially contain concurrent GPU kernel execution.
We implement an alternate token technology operate that interleaves two CUDA streams, working the next operations iteratively:
Program stream ipercent2 to: (A) look forward to stream (i-1)%2 to finish its technology of token i-1, (B) use the up to date tensors to calculate the token i, (C) run the EOS take a look at for token i on the GPU, and (D) carry out a (non-blocking) copy of the EOS take a look at outcome to pinned reminiscence on the CPU.
On the default CUDA stream, look forward to stream (i-1)%2 to finish its technology of token i-1.
On the default CUDA stream, examine if the stopping standards for token i-1 had been met. If that’s the case, halt the generator and return. In any other case, increment i and return to step 1.
Whereas beforehand, the initialization of token i technology was blocked by the EOS take a look at on token i-1, the usage of CUDA streams permits us to program the technology of token i earlier than we examine the results of the EOS take a look at on token i-1. In observe, the EOS take a look at for token i-1 on the CPU runs whereas the GPU is computing token i.
@torch.inference_mode()
def generate_sequence_pipelined(
mannequin,
max_seqlen,
batch_size,
use_cache=False,
use_static_cache=False
):
# Initialize prompts with BOS token
all_tokens = torch.full(
(batch_size, 1),
config.bos_token_id,
machine=DEVICE,
dtype=torch.lengthy
)
completed = torch.zeros(batch_size, machine=DEVICE, dtype=torch.bool)
past_key_values = None
# Initialize static cache if requested
if use_cache and use_static_cache:
past_key_values = StaticCache(
config=config,
max_batch_size=batch_size,
max_cache_len=max_seqlen,
machine=DEVICE,
dtype=mannequin.dtype
)
# Initialize cache place monitoring for static cache
cache_positions = torch.arange(max_seqlen, machine=DEVICE)
# Twin streams for pipelining
streams = [torch.cuda.Stream(), torch.cuda.Stream()]
stop_host = [
torch.tensor(False, pin_memory=True),
torch.tensor(False, pin_memory=True)
]
for i in vary(max_seqlen):
curr_idx, prev_idx = i % 2, (i+1) % 2
curr_s, prev_s = streams[curr_idx], streams[prev_idx]
# Launch iteration i in present stream
with torch.cuda.stream(curr_s):
# program stream to attend for earlier stream to finish
curr_s.wait_stream(prev_s)
current_input = (
all_tokens if past_key_values is None
else all_tokens[:, -1:]
)
cache_position = (
cache_positions[i:i+1] if use_static_cache else None
)
outputs = mannequin(
current_input,
past_key_values=past_key_values,
cache_position=cache_position,
use_cache=use_cache
)
past_key_values = outputs.past_key_values
logits = outputs.logits[:, -1, :]
new_tokens = torch.argmax(logits, dim=-1)
all_tokens = torch.cat(
[all_tokens, new_tokens.unsqueeze(-1)],
dim=-1
)
completed |= (new_tokens == config.eos_token_id)
stop_gpu = torch.all(completed)
stop_host[curr_idx].copy_(stop_gpu, non_blocking=True)
# Test earlier iteration's cease sign
torch.cuda.current_stream().wait_stream(prev_s)
if stop_host[prev_idx].merchandise():
print(f"All sequences completed at step {i}")
break
return all_tokensThe picture beneath captures the nsys hint for our new token generator:
Within the CUDA part of the hint we are able to see the usage of two CUDA streams, with token technology being handed forwards and backwards in a type of ping-pong impact: One stream generates all the odd tokens and second all the even tokens. The CPU is about half a step forward of the GPU — permitting it to program step i whereas the GPU is computing step i-1. The CPU-side EOS stop-check of step i-1 (in crimson) happens after step i is absolutely programmed (and has began working). Most significantly, we now discover the GPU utilization to be constant — the idling we noticed earlier than is gone.
The CUDA stream interleaving leads to an extra efficiency increase, as proven within the desk beneath:
We’d anticipate the advantage of the ping-pong resolution we’ve got carried out to be impacted by the ratio between the GPU idle time (i.e., the overhead of kernel loading) and the kernel computation time. To check this, we repair the sequence size at 100 and rerun the benchmark for various batch sizes:
As anticipated, the best efficiency acquire, 11.6%, happens when the batch dimension is smallest and the kernel computation load is at its lowest. Because the kernel compute will increase, the ratio of kernel loading to kernel compute time decreases as does the influence of CUDA stream interleaving.
Observe that there’s some overhead to the usage of CUDA streams. This may be demonstrated by evaluating our interleaving resolution to a token generator that skips the EOS take a look at altogether:
The Potential Efficiency Pitfalls of Utilizing CUDA Streams
CUDA streams ought to be used with excessive warning. When utilizing the default stream we are able to depend on PyTorch to carry out any vital synchronization when information is moved round. Nonetheless, when utilizing CUDA streams, we should guarantee acceptable synchronization explicitly. Particularly, we should guarantee acceptable information switch between the streams. In any other case, we might expertise CUDA errors (e.g., “device-side assert triggered”) — if we’re fortunate. If we’re much less fortunate, we might expertise information corruption with out even figuring out it. See the PyTorch CUDA stream documentation for extra particulars on acceptable use.
For AI/ML workloads with massive CUDA reminiscence utilization, akin to LLMs, one other consideration is reminiscence utilization. The PyTorch caching allocator manages reminiscence on a per-stream foundation; utilizing a number of streams can result in elevated reminiscence reservation and fragmentation. These may lead to elevated reminiscence faults which may overshadow the potential positive aspects from the usage of streams.
Outcomes
Within the desk beneath we summarize the runtime outcomes of making use of static caching, compilation, and pipelining on a batch of 32 sequences and a most sequence size of 100. The outcomes are sorted in rising order of efficiency:
Within the case of our toy GPT-2 mannequin, the very best outcomes — practically 5 instances the baseline efficiency — are achieved when using PyTorch compilation and the CUDA stream interleaving technique mentioned on this publish. Nonetheless, as we’ve got seen, the influence of CUDA interleaving may range vastly primarily based on the properties of the workload and runtime setting, notably on the ratio between the kernel loading time and the kernel compute time. Please be sure you run your personal benchmarks earlier than adopting this technique.
Abstract
In high-performance AI engineering, any trace of GPU under-utilization presents a chance for optimization. One of many major optimization instruments on NVIDIA GPUs is CUDA streams. On this publish, we demonstrated their use in fixing the idle GPU time that outcomes from the host-device synchronization related to early-stopping in PyTorch-native autoregressive token technology. By interleaving CUDA streams in a “ping-pong” sample, we efficiently hid the latency imposed by the EOS-check which resulted in a significant enhance the workload’s throughput. By combining this system with the well-known strategies of mannequin compilation and static caching, we are able to maximize the efficiency of PyTorch-native inference.
