which have pervaded almost each side of our each day lives are autoregressive decoder fashions. These fashions apply compute-heavy kernel operations to churn out tokens one after the other in a fashion that, at first look, appears extraordinarily inefficient. Given the big demand for generative AI, it’s no shock that extraordinary engineering effort is being invested into its optimization. Whether or not it’s by customized CUDA kernels, CUDA Graphs, devoted AI accelerators, or speculative sampling — any method that reduces latency and/or value by even a fraction of a proportion is a win.
On this submit, we show a way for optimizing token era in PyTorch utilizing CUDA stream interleaving. Whereas easy to implement, the tactic addresses a selected, usually ignored bottleneck and may result in significant efficiency boosts. Whereas pipelining mannequin execution utilizing CUDA streams is widespread in AI methods engineering, we didn’t discover any tutorial documenting the particular PyTorch-level software we describe right here. Should you discover the method helpful, please be so variety as to reference this submit.
To facilitate our dialogue, we are going to use a easy GPT-2 PyTorch decoder mannequin from HuggingFace’s transformers (v5.1.0) library. We are going to run our experiments on an NVIDIA L40S GPU and PyTorch (2.10.0).
Disclaimer: The code we are going to 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 methodology we are going to talk about can fluctuate enormously primarily based on the main points of your mannequin and runtime setting. Please you should definitely run your individual benchmarks earlier than integrating its use.
Our focus on this submit is on PyTorch-native inference workloads which stay extraordinarily prevalent in improvement and check settings. Nevertheless, it is very important be aware that for manufacturing environments devoted LLM inference libraries corresponding to vLLM or NVIDIA TensorRT-LLM are likely to ship better efficiency and needs to be used every time related.
A Toy GPT-2 Mannequin
To simplify our dialogue, we are going to 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 era 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,
gadget=DEVICE,
dtype=torch.lengthy
)
completed = torch.zeros(batch_size, gadget=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 era with batch measurement {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 under we seize the outcomes for a batch measurement of 32 and a number of other totally different sequence lengths:
Because the sequence size doubles, the runtime quadruples — showing to comply with a basic 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 tackle 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 era 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 era 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,
gadget=DEVICE,
dtype=torch.lengthy
)
completed = torch.zeros(batch_size, gadget=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 deal with this we discover two strategies, expandable reminiscence allocations and static KV caching.
Expandable CUDA Reminiscence Allocations
To cut 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 needs 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 in the course of the era progresses. HuggingFace helps a fixed-size cache, StaticCache, which pre-allocates a most cache measurement for the KV pairs and reduces pressure on the CUDA reminiscence allocator. The drawback of utilizing StaticCache is that the total size of the cache participates within the consideration computation at every token era 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 can be run on full 400X400-sized tensors.
Within the code block under we improve our sequence generator to assist using 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,
gadget=DEVICE,
dtype=torch.lengthy
)
completed = torch.zeros(batch_size, gadget=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,
gadget=DEVICE,
dtype=mannequin.dtype
)
else:
past_key_values = None
# Initialize cache place monitoring for static cache
cache_positions = torch.arange(max_seqlen, gadget=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 under:
Utilizing a fixed-sized cache enormously improves reminiscence utilization as indicated by the lower in reminiscence fragmentation. Nevertheless, its affect on runtime efficiency is blended — 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 submit, Optimizing Transformer Models for Variable-Length Input Sequences, we coated some PyTorch strategies for computing consideration sparsely to scale back computation waste. For manufacturing settings, libraries corresponding to vLLM use PagedAttention for maximizing reminiscence utilization. These strategies are exterior the scope of this submit.
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 under:
Word that we will apply mannequin compilation when utilizing dynamic caching, as nicely. Nevertheless, torch.compile supplies the perfect outcomes when the computation graph consists of fixed-sized tensors (e.g., see here for extra particulars).
The Efficiency Penalty of Early Stopping
An integral a part of widespread token mills is checking for the end-of-sequence (EOS) on the finish of every step. With out this check, token mills would at all times run for max_seqlen, even when all of the sequences within the batch have ended. This might end in appreciable computation waste and pointless latency — particularly when widespread 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 era. 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, to be able to consider the conditional if assertion, the CPU should look forward to the GPU to finish its computation and duplicate the contents of the tensor to host reminiscence. Whereas the CPU waits, it’s blocked from executing the subsequent step of the token era loop, or extra precisely, it’s blocked from loading the subsequent computation kernels onto the GPU.
To measure the affect 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 recent post 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,
gadget=DEVICE,
dtype=torch.lengthy
)
completed = torch.zeros(batch_size, gadget=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,
gadget=DEVICE,
dtype=mannequin.dtype
)
else:
past_key_values = None
# Initialize cache place monitoring for static cache
cache_positions = torch.arange(max_seqlen, gadget=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}", shade="blue"):
with nvtx.annotate("Mannequin Ahead", shade="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("Verify Cease Situation", shade="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 begin 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 prepare.pyThe next hint, captured for a batch measurement 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. This can be a direct results of the synchronization occasion triggered by the EOS check.
In production-grade implementations such synchronization points are prevented by some mixture of 1) use of decrease stage (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) constantly and asynchronously making ready subsequent requests whereas the EOS test is in progress.
Within the subsequent part, we show a way 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 so forth.) 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., here and here) we demonstrated using CUDA streams in pipelining widespread AI/ML workloads, e.g., executing a mannequin on batch N whereas making ready batch N+1. On this submit we are going to 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 is not going to essentially contain concurrent GPU kernel execution.
We implement another token era operate that interleaves two CUDA streams, operating the next operations iteratively:
Program stream ipercent2 to: (A) look forward to stream (i-1)%2 to finish its era of token i-1, (B) use the up to date tensors to calculate the token i, (C) run the EOS check for token i on the GPU, and (D) carry out a (non-blocking) copy of the EOS check consequence to pinned reminiscence on the CPU.
On the default CUDA stream, look forward to stream (i-1)%2 to finish its era of token i-1.
On the default CUDA stream, test if the stopping standards for token i-1 have 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 era was blocked by the EOS check on token i-1, using CUDA streams permits us to program the era of token i earlier than we test the results of the EOS check on token i-1. In follow, the EOS check 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,
gadget=DEVICE,
dtype=torch.lengthy
)
completed = torch.zeros(batch_size, gadget=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,
gadget=DEVICE,
dtype=mannequin.dtype
)
# Initialize cache place monitoring for static cache
cache_positions = torch.arange(max_seqlen, gadget=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)
# Verify 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 under captures the nsys hint for our new token generator:
Within the CUDA part of the hint we will see using two CUDA streams, with token era being handed forwards and backwards in a form of ping-pong impact: One stream generates the entire odd tokens and second the entire 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 totally programmed (and has began operating). 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 under:
We might anticipate the advantage of the ping-pong resolution now we have 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 a lot of batch sizes:
As anticipated, the best efficiency acquire, 11.6%, happens when the batch measurement 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 affect of CUDA stream interleaving.
Word that there’s some overhead to using CUDA streams. This may be demonstrated by evaluating our interleaving resolution to a token generator that skips the EOS check altogether:
The Potential Efficiency Pitfalls of Utilizing CUDA Streams
CUDA streams needs to be used with excessive warning. When utilizing the default stream we will depend on PyTorch to carry out any needed synchronization when information is moved round. Nevertheless, when utilizing CUDA streams, we should guarantee applicable synchronization explicitly. Specifically, we should guarantee applicable information switch between the streams. In any other case, we could expertise CUDA errors (e.g., “device-side assert triggered”) — if we’re fortunate. If we’re much less fortunate, we could expertise information corruption with out even understanding it. See the PyTorch CUDA stream documentation for extra particulars on applicable use.
For AI/ML workloads with massive CUDA reminiscence utilization, corresponding 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 might end in elevated reminiscence faults that may overshadow the potential good points from using streams.
Outcomes
Within the desk under 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 perfect outcomes — almost 5 instances the baseline efficiency — are achieved when using PyTorch compilation and the CUDA stream interleaving methodology mentioned on this submit. Nevertheless, as now we have seen, the affect of CUDA interleaving might fluctuate enormously 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 you should definitely run your individual benchmarks earlier than adopting this methodology.
Abstract
In high-performance AI engineering, any trace of GPU under-utilization presents a chance for optimization. One of many main optimization instruments on NVIDIA GPUs is CUDA streams. On this submit, 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 era. By interleaving CUDA streams in a “ping-pong” sample, we efficiently hid the latency imposed by the EOS-check which resulted in a significant improve the workload’s throughput. By combining this system with the well-known strategies of mannequin compilation and static caching, we will maximize the efficiency of PyTorch-native inference.
