Skip to content

Latest commit

 

History

History
515 lines (446 loc) · 28.5 KB

README_GAUDI.md

File metadata and controls

515 lines (446 loc) · 28.5 KB
Raw

vLLM with Intel® Gaudi® AI Accelerators

This README provides instructions on running vLLM with Intel Gaudi devices.

Requirements and Installation

Please follow the instructions provided in the Gaudi Installation Guide to set up the environment. To achieve the best performance, please follow the methods outlined in the Optimizing Training Platform Guide.

Requirements

  • OS: Ubuntu 22.04 LTS
  • Python: 3.10
  • Intel Gaudi accelerator
  • Intel Gaudi software version 1.17.0

To verify that the Intel Gaudi software was correctly installed, run:

$ hl-smi # verify that hl-smi is in your PATH and each Gaudi accelerator is visible
$ apt list --installed | grep habana # verify that habanalabs-firmware-tools, habanalabs-graph, habanalabs-rdma-core and habanalabs-thunk are installed
$ pip list | grep habana # verify that habana-torch-plugin, habana-torch-dataloader, habana-pyhlml and habana-media-loader are installed
$ pip list | grep neural # verify that neural-compressor is installed

Refer to Intel Gaudi Software Stack Verification for more details.

Run Docker Image

It is highly recommended to use the latest Docker image from Intel Gaudi vault. Refer to the Intel Gaudi documentation for more details.

Use the following commands to run a Docker image:

$ docker pull vault.habana.ai/gaudi-docker/1.17.0/ubuntu22.04/habanalabs/pytorch-installer-2.3.1:latest
$ docker run -it --runtime=habana -e HABANA_VISIBLE_DEVICES=all -e OMPI_MCA_btl_vader_single_copy_mechanism=none --cap-add=sys_nice --net=host --ipc=host vault.habana.ai/gaudi-docker/1.17.0/ubuntu22.04/habanalabs/pytorch-installer-2.3.1:latest

Build and Install vLLM

Currently, the latest features and performance optimizations are developed in Gaudi's vLLM-fork and we periodically upstream them to vLLM main repo. To install latest HabanaAI/vLLM-fork, run the following:

$ git clone https://github.com/HabanaAI/vllm-fork.git
$ cd vllm-fork
$ git checkout habana_main
$ pip install -e .

Supported Features

  • Offline batched inference
  • Online inference via OpenAI-Compatible Server
  • HPU autodetection - no need to manually select device within vLLM
  • Paged KV cache with algorithms enabled for Intel Gaudi accelerators
  • Custom Intel Gaudi implementations of Paged Attention, KV cache ops, prefill attention, Root Mean Square Layer Normalization, Rotary Positional Encoding
  • Tensor parallelism support for multi-card inference
  • Inference with HPU Graphs for accelerating low-batch latency and throughput
  • Attention with Linear Biases (ALiBi)
  • INC quantization

Unsupported Features

  • Beam search
  • LoRA adapters
  • AWQ quantization
  • Prefill chunking (mixed-batch inferencing)

Supported Configurations

The following configurations have been validated to be function with Gaudi2 devices. Configurations that are not listed may or may not work.

Performance Tuning

Execution modes

Currently in vLLM for HPU we support four execution modes, depending on selected HPU PyTorch Bridge backend (via PT_HPU_LAZY_MODE environment variable), and --enforce-eager flag.

PT_HPU_LAZY_MODE enforce_eager execution mode
0 0 torch.compile
0 1 PyTorch eager mode
1 0 HPU Graphs
1 1 PyTorch lazy mode

Warning

In 1.17.0, all modes utilizing PT_HPU_LAZY_MODE=0 are highly experimental and should be only used for validating functional correctness. Their performance will be improved in the next releases. For obtaining the best performance in 1.17.0, please use HPU Graphs, or PyTorch lazy mode.

Bucketing mechanism

Intel Gaudi accelerators work best when operating on models with fixed tensor shapes. Intel Gaudi Graph Compiler is responsible for generating optimized binary code that implements the given model topology on Gaudi. In its default configuration, the produced binary code may be heavily dependent on input and output tensor shapes, and can require graph recompilation when encountering differently shaped tensors within the same topology. While the resulting binaries utilize Gaudi efficiently, the compilation itself may introduce a noticeable overhead in end-to-end execution. In a dynamic inference serving scenario, there is a need to minimize the number of graph compilations and reduce the risk of graph compilation occurring during server runtime. Currently it is achieved by "bucketing" model's forward pass across two dimensions - batch_size and sequence_length.

Note

Bucketing allows us to reduce the number of required graphs significantly, but it does not handle any graph compilation and device code generation - this is done in warmup and HPUGraph capture phase.

Bucketing ranges are determined with 3 parameters - min, step and max. They can be set separately for prompt and decode phase, and for batch size and sequence length dimension. These parameters can be observed in logs during vLLM startup:

INFO 08-01 21:37:59 hpu_model_runner.py:493] Prompt bucket config (min, step, max_warmup) bs:[1, 32, 4], seq:[128, 128, 1024]
INFO 08-01 21:37:59 hpu_model_runner.py:499] Generated 24 prompt buckets: [(1, 128), (1, 256), (1, 384), (1, 512), (1, 640), (1, 768), (1, 896), (1, 1024), (2, 128), (2, 256), (2, 384), (2, 512), (2, 640), (2, 768), (2, 896), (2, 1024), (4, 128), (4, 256), (4, 384), (4, 512), (4, 640), (4, 768), (4, 896), (4, 1024)]
INFO 08-01 21:37:59 hpu_model_runner.py:504] Decode bucket config (min, step, max_warmup) bs:[1, 128, 4], seq:[128, 128, 2048]
INFO 08-01 21:37:59 hpu_model_runner.py:509] Generated 48 decode buckets: [(1, 128), (1, 256), (1, 384), (1, 512), (1, 640), (1, 768), (1, 896), (1, 1024), (1, 1152), (1, 1280), (1, 1408), (1, 1536), (1, 1664), (1, 1792), (1, 1920), (1, 2048), (2, 128), (2, 256), (2, 384), (2, 512), (2, 640), (2, 768), (2, 896), (2, 1024), (2, 1152), (2, 1280), (2, 1408), (2, 1536), (2, 1664), (2, 1792), (2, 1920), (2, 2048), (4, 128), (4, 256), (4, 384), (4, 512), (4, 640), (4, 768), (4, 896), (4, 1024), (4, 1152), (4, 1280), (4, 1408), (4, 1536), (4, 1664), (4, 1792), (4, 1920), (4, 2048)]

min determines the lowest value of the bucket. step determines the interval between buckets, and max determines the upper bound of the bucket. Furthermore, interval between min and step has special handling - min gets multiplied by consecutive powers of two, until step gets reached. We call this the ramp-up phase and it is used for handling lower batch sizes with minimum wastage, while allowing larger padding on larger batch sizes.

Example (with ramp-up)

min = 2, step = 32, max = 64
=> ramp_up = (2, 4, 8, 16)
=> stable = (32, 64)
=> buckets = ramp_up + stable => (2, 4, 8, 16, 32, 64)

Example (without ramp-up)

min = 128, step = 128, max = 512
=> ramp_up = ()
=> stable = (128, 256, 384, 512)
=> buckets = ramp_up + stable => (128, 256, 384, 512)

In the logged scenario, 24 buckets were generated for prompt (prefill) runs, and 48 buckets for decode runs. Each bucket corresponds to a separate optimized device binary for a given model with specified tensor shapes. Whenever a batch of requests is processed, it is padded across batch and sequence length dimension to the smallest possible bucket.

Warning

If a request exceeds maximum bucket size in any dimension, it will be processed without padding, and its processing may require a graph compilation, potentially significantly increasing end-to-end latency. The boundaries of the buckets are user-configurable via environment variables, and upper bucket boundaries can be increased to avoid such scenario.

As an example, if a request of 3 sequences, with max sequence length of 412 comes in to an idle vLLM server, it will be padded executed as (4, 512) prefill bucket, as batch_size (number of sequences) will be padded to 4 (closest batch_size dimension higher than 3), and max sequence length will be padded to 512 (closest sequence length dimension higher than 412). After prefill stage, it will be executed as (4, 512) decode bucket and will continue as that bucket until either batch dimension changes (due to request being finished) - in which case it will become a (2, 512) bucket, or context length increases above 512 tokens, in which case it will become (4, 640) bucket.

Note

Bucketing is transparent to a client - padding in sequence length dimension is never returned to the client, and padding in batch dimension does not create new requests.

Warmup

Warmup is an optional, but highly recommended step occurring before vLLM server starts listening. It executes a forward pass for each bucket with dummy data. The goal is to pre-compile all graphs and not incur any graph compilation overheads within bucket boundaries during server runtime. Each warmup step is logged during vLLM startup:

INFO 08-01 22:26:47 hpu_model_runner.py:1066] [Warmup][Prompt][1/24] batch_size:4 seq_len:1024 free_mem:79.16 GiB
INFO 08-01 22:26:47 hpu_model_runner.py:1066] [Warmup][Prompt][2/24] batch_size:4 seq_len:896 free_mem:55.43 GiB
INFO 08-01 22:26:48 hpu_model_runner.py:1066] [Warmup][Prompt][3/24] batch_size:4 seq_len:768 free_mem:55.43 GiB
...
INFO 08-01 22:26:59 hpu_model_runner.py:1066] [Warmup][Prompt][24/24] batch_size:1 seq_len:128 free_mem:55.43 GiB
INFO 08-01 22:27:00 hpu_model_runner.py:1066] [Warmup][Decode][1/48] batch_size:4 seq_len:2048 free_mem:55.43 GiB
INFO 08-01 22:27:00 hpu_model_runner.py:1066] [Warmup][Decode][2/48] batch_size:4 seq_len:1920 free_mem:55.43 GiB
INFO 08-01 22:27:01 hpu_model_runner.py:1066] [Warmup][Decode][3/48] batch_size:4 seq_len:1792 free_mem:55.43 GiB
...
INFO 08-01 22:27:16 hpu_model_runner.py:1066] [Warmup][Decode][47/48] batch_size:2 seq_len:128 free_mem:55.43 GiB
INFO 08-01 22:27:16 hpu_model_runner.py:1066] [Warmup][Decode][48/48] batch_size:1 seq_len:128 free_mem:55.43 GiB

This example uses the same buckets as in Bucketing mechanism section. Each output line corresponds to execution of a single bucket. When bucket is executed for the first time, its graph is compiled and can be reused later on, skipping further graph compilations.

Tip

Compiling all the buckets might take some time and can be turned off with VLLM_SKIP_WARMUP=true environment variable. Keep in mind that if you do that, you may face graph compilations once executing a given bucket for the first time. It is fine to disable warmup for development, but it's highly recommended to enable it in deployment.

HPU Graph capture

HPU Graphs are currently the most performant execution method of vLLM on Intel Gaudi. When HPU Graphs are enabled, execution graphs will be traced (recorded) ahead of time (after performing warmup), to be later replayed during inference, significantly reducing host overheads. Recording can take large amounts of memory, which needs to be taken into account when allocating KV cache. Enabling HPU Graphs will impact the number of available KV cache blocks, but vLLM provides user-configurable variables to control memory management.

When HPU Graphs are being used, they share the common memory pool ("usable memory") as KV cache, determined by gpu_memory_utilization flag (0.9 by default). Before KV cache gets allocated, model weights are loaded onto the device, and a forward pass of the model is executed on dummy data, to estimate memory usage. Only after that, gpu_memory_utilization flag is utilized - at its default value, will mark 90% of free device memory at that point as usable. Next, KV cache gets allocated, model is warmed up, and HPU Graphs are captured. Environment variable VLLM_GRAPH_RESERVED_MEM defines the ratio of memory reserved for HPU Graphs capture. With its default value (VLLM_GRAPH_RESERVED_MEM=0.1), 10% of usable memory will be reserved for graph capture (later referred to as "usable graph memory"), and the remaining 90% will be utilized for KV cache. Environment variable VLLM_GRAPH_PROMPT_RATIO determines the ratio of usable graph memory reserved for prefill and decode graphs. By default (VLLM_GRAPH_PROMPT_RATIO=0.3), both stages have equal memory constraints. Lower value corresponds to less usable graph memory reserved for prefill stage, e.g. VLLM_GRAPH_PROMPT_RATIO=0.2 will reserve 20% of usable graph memory for prefill graphs, and 80% of usable graph memory for decode graphs.

Note

gpu_memory_utilization does not correspond to the absolute memory usage across HPU. It specifies the memory margin after loading the model and performing a profile run. If device has 100 GiB of total memory, and 50 GiB of free memory after loading model weights and executing profiling run, gpu_memory_utilization at its default value will mark 90% of 50 GiB as usable, leaving 5 GiB of margin, regardless of total device memory.

User can also configure the strategy for capturing HPU Graphs for prompt and decode stages separately. Strategy affects the order of capturing graphs. There are two strategies implemented: - max_bs - graph capture queue will sorted in descending order by their batch sizes. Buckets with equal batch sizes are sorted by sequence length in ascending order (e.g. (64, 128), (64, 256), (32, 128), (32, 256), (1, 128), (1,256)), default strategy for decode - min_tokens - graph capture queue will be sorted in ascending order by the number of tokens each graph processes (batch_size*sequence_length), default strategy for prompt

When there's large amount of requests pending, vLLM scheduler will attempt to fill the maximum batch size for decode as soon as possible. When a request is finished, decode batch size decreases. When that happens, vLLM will attempt to schedule a prefill iteration for requests in the waiting queue, to fill the decode batch size to its previous state. This means that in a full load scenario, decode batch size is often at its maximum, which makes large batch size HPU Graphs crucial to capture, as reflected by max_bs strategy. On the other hand, prefills will be executed most frequently with very low batch sizes (1-4), which is reflected in min_tokens strategy.

Note

VLLM_GRAPH_PROMPT_RATIO does not set a hard limit on memory taken by graphs for each stage (prefill and decode). vLLM will first attempt to use up entirety of usable prefill graph memory (usable graph memory * VLLM_GRAPH_PROMPT_RATIO) for capturing prefill HPU Graphs, next it will attempt do the same for decode graphs and usable decode graph memory pool. If one stage is fully captured, and there is unused memory left within usable graph memory pool, vLLM will attempt further graph capture for the other stage, until no more HPU Graphs can be captured without exceeding reserved memory pool. The behavior on that mechanism can be observed in the example below.

Each described step is logged by vLLM server, as follows (negative values correspond to memory being released):

INFO 08-02 17:37:44 hpu_model_runner.py:493] Prompt bucket config (min, step, max_warmup) bs:[1, 32, 4], seq:[128, 128, 1024]
INFO 08-02 17:37:44 hpu_model_runner.py:499] Generated 24 prompt buckets: [(1, 128), (1, 256), (1, 384), (1, 512), (1, 640), (1, 768), (1, 896), (1, 1024), (2, 128), (2, 256), (2, 384), (2, 512), (2, 640), (2, 768), (2, 896), (2, 1024), (4, 128), (4, 256), (4, 384), (4, 512), (4, 640), (4, 768), (4, 896), (4, 1024)]
INFO 08-02 17:37:44 hpu_model_runner.py:504] Decode bucket config (min, step, max_warmup) bs:[1, 128, 4], seq:[128, 128, 2048]
INFO 08-02 17:37:44 hpu_model_runner.py:509] Generated 48 decode buckets: [(1, 128), (1, 256), (1, 384), (1, 512), (1, 640), (1, 768), (1, 896), (1, 1024), (1, 1152), (1, 1280), (1, 1408), (1, 1536), (1, 1664), (1, 1792), (1, 1920), (1, 2048), (2, 128), (2, 256), (2, 384), (2, 512), (2, 640), (2, 768), (2, 896), (2, 1024), (2, 1152), (2, 1280), (2, 1408), (2, 1536), (2, 1664), (2, 1792), (2, 1920), (2, 2048), (4, 128), (4, 256), (4, 384), (4, 512), (4, 640), (4, 768), (4, 896), (4, 1024), (4, 1152), (4, 1280), (4, 1408), (4, 1536), (4, 1664), (4, 1792), (4, 1920), (4, 2048)]
INFO 08-02 17:37:52 hpu_model_runner.py:430] Pre-loading model weights on hpu:0 took 14.97 GiB of device memory (14.97 GiB/94.62 GiB used) and 2.95 GiB of host memory (475.2 GiB/1007 GiB used)
INFO 08-02 17:37:52 hpu_model_runner.py:438] Wrapping in HPU Graph took 0 B of device memory (14.97 GiB/94.62 GiB used) and -252 KiB of host memory (475.2 GiB/1007 GiB used)
INFO 08-02 17:37:52 hpu_model_runner.py:442] Loading model weights took in total 14.97 GiB of device memory (14.97 GiB/94.62 GiB used) and 2.95 GiB of host memory (475.2 GiB/1007 GiB used)
INFO 08-02 17:37:54 hpu_worker.py:134] Model profiling run took 504 MiB of device memory (15.46 GiB/94.62 GiB used) and 180.9 MiB of host memory (475.4 GiB/1007 GiB used)
INFO 08-02 17:37:54 hpu_worker.py:158] Free device memory: 79.16 GiB, 39.58 GiB usable (gpu_memory_utilization=0.5), 15.83 GiB reserved for HPUGraphs (VLLM_GRAPH_RESERVED_MEM=0.4), 23.75 GiB reserved for KV cache
INFO 08-02 17:37:54 hpu_executor.py:85] # HPU blocks: 1519, # CPU blocks: 0
INFO 08-02 17:37:54 hpu_worker.py:190] Initializing cache engine took 23.73 GiB of device memory (39.2 GiB/94.62 GiB used) and -1.238 MiB of host memory (475.4 GiB/1007 GiB used)
INFO 08-02 17:37:54 hpu_model_runner.py:1066] [Warmup][Prompt][1/24] batch_size:4 seq_len:1024 free_mem:55.43 GiB
...
INFO 08-02 17:38:22 hpu_model_runner.py:1066] [Warmup][Decode][48/48] batch_size:1 seq_len:128 free_mem:55.43 GiB
INFO 08-02 17:38:22 hpu_model_runner.py:1159] Using 15.85 GiB/55.43 GiB of free device memory for HPUGraphs, 4.755 GiB for prompt and 11.095 GiB for decode (VLLM_GRAPH_PROMPT_RATIO=0.3)
INFO 08-02 17:38:22 hpu_model_runner.py:1066] [Warmup][Graph/Prompt][1/24] batch_size:1 seq_len:128 free_mem:55.43 GiB
...
INFO 08-02 17:38:26 hpu_model_runner.py:1066] [Warmup][Graph/Prompt][11/24] batch_size:1 seq_len:896 free_mem:48.77 GiB
INFO 08-02 17:38:27 hpu_model_runner.py:1066] [Warmup][Graph/Decode][1/48] batch_size:4 seq_len:128 free_mem:47.51 GiB
...
INFO 08-02 17:38:41 hpu_model_runner.py:1066] [Warmup][Graph/Decode][48/48] batch_size:1 seq_len:2048 free_mem:47.35 GiB
INFO 08-02 17:38:41 hpu_model_runner.py:1066] [Warmup][Graph/Prompt][12/24] batch_size:4 seq_len:256 free_mem:47.35 GiB
INFO 08-02 17:38:42 hpu_model_runner.py:1066] [Warmup][Graph/Prompt][13/24] batch_size:2 seq_len:512 free_mem:45.91 GiB
INFO 08-02 17:38:42 hpu_model_runner.py:1066] [Warmup][Graph/Prompt][14/24] batch_size:1 seq_len:1024 free_mem:44.48 GiB
INFO 08-02 17:38:43 hpu_model_runner.py:1066] [Warmup][Graph/Prompt][15/24] batch_size:2 seq_len:640 free_mem:43.03 GiB
INFO 08-02 17:38:43 hpu_model_runner.py:1128] Graph/Prompt captured:15 (62.5%) used_mem:14.03 GiB buckets:[(1, 128), (1, 256), (1, 384), (1, 512), (1, 640), (1, 768), (1, 896), (1, 1024), (2, 128), (2, 256), (2, 384), (2, 512), (2, 640), (4, 128), (4, 256)]
INFO 08-02 17:38:43 hpu_model_runner.py:1128] Graph/Decode captured:48 (100.0%) used_mem:161.9 MiB buckets:[(1, 128), (1, 256), (1, 384), (1, 512), (1, 640), (1, 768), (1, 896), (1, 1024), (1, 1152), (1, 1280), (1, 1408), (1, 1536), (1, 1664), (1, 1792), (1, 1920), (1, 2048), (2, 128), (2, 256), (2, 384), (2, 512), (2, 640), (2, 768), (2, 896), (2, 1024), (2, 1152), (2, 1280), (2, 1408), (2, 1536), (2, 1664), (2, 1792), (2, 1920), (2, 2048), (4, 128), (4, 256), (4, 384), (4, 512), (4, 640), (4, 768), (4, 896), (4, 1024), (4, 1152), (4, 1280), (4, 1408), (4, 1536), (4, 1664), (4, 1792), (4, 1920), (4, 2048)]
INFO 08-02 17:38:43 hpu_model_runner.py:1206] Warmup finished in 49 secs, allocated 14.19 GiB of device memory
INFO 08-02 17:38:43 hpu_executor.py:91] init_cache_engine took 37.92 GiB of device memory (53.39 GiB/94.62 GiB used) and 57.86 MiB of host memory (475.4 GiB/1007 GiB used)

Recommended vLLM Parameters

  • We recommend running inference on Gaudi 2 with block_size of 128 for BF16 data type. Using default values (16, 32) might lead to sub-optimal performance due to Matrix Multiplication Engine under-utilization (see Gaudi Architecture).
  • For max throughput on Llama 7B, we recommend running with batch size of 128 or 256 and max context length of 2048 with HPU Graphs enabled. If you encounter out-of-memory issues, see troubleshooting section.

Environment variables

Diagnostic and profiling knobs:

  • VLLM_PROFILER_ENABLED: if true, high level profiler will be enabled. Resulting JSON traces can be viewed in perfetto.habana.ai. Disabled by default.
  • VLLM_HPU_LOG_STEP_GRAPH_COMPILATION: if true, will log graph compilations per each vLLM engine step, only when there was any - highly recommended to use alongside PT_HPU_METRICS_GC_DETAILS=1. Disabled by default.
  • VLLM_HPU_LOG_STEP_GRAPH_COMPILATION_ALL: if true, will log graph compilations per each vLLM engine step, always, even if there were none. Disabled by default.
  • VLLM_HPU_LOG_STEP_CPU_FALLBACKS: if true, will log cpu fallbacks per each vLLM engine step, only when there was any. Disabled by default.
  • VLLM_HPU_LOG_STEP_CPU_FALLBACKS_ALL: if true, will log cpu fallbacks per each vLLM engine step, always, even if there were none. Disabled by default.

Performance tuning knobs:

  • VLLM_SKIP_WARMUP: if true, warmup will be skipped, false by default
  • VLLM_GRAPH_RESERVED_MEM: percentage of memory dedicated for HPUGraph capture, 0.1 by default
  • VLLM_GRAPH_PROMPT_RATIO: percentage of reserved graph memory dedicated for prompt graphs, 0.3 by default
  • VLLM_GRAPH_PROMPT_STRATEGY: strategy determining order of prompt graph capture, min_tokens or max_bs, min_tokens by default
  • VLLM_GRAPH_DECODE_STRATEGY: strategy determining order of decode graph capture, min_tokens or max_bs, max_bs by default
  • VLLM_{phase}_{dim}_BUCKET_{param} - collection of 12 environment variables configuring ranges of bucketing mechanism
    • {phase} is either PROMPT or DECODE
    • {dim} is either BS, SEQ or BLOCK
    • {param} is either MIN, STEP or MAX
    • Default values:

      • Prompt:

        • batch size min (VLLM_PROMPT_BS_BUCKET_MIN): 1
        • batch size step (VLLM_PROMPT_BS_BUCKET_STEP): min(max_num_seqs, 32)
        • batch size max (VLLM_PROMPT_BS_BUCKET_MAX): min(max_num_seqs, 64)
        • sequence length min (VLLM_PROMPT_SEQ_BUCKET_MIN): block_size
        • sequence length step (VLLM_PROMPT_SEQ_BUCKET_STEP): block_size
        • sequence length max (VLLM_PROMPT_SEQ_BUCKET_MAX): max_model_len

      • Decode:

        • batch size min (VLLM_DECODE_BS_BUCKET_MIN): 1
        • batch size step (VLLM_DECODE_BS_BUCKET_STEP): min(max_num_seqs, 32)
        • batch size max (VLLM_DECODE_BS_BUCKET_MAX): max_num_seqs
        • block size min (VLLM_DECODE_BLOCK_BUCKET_MIN): block_size
        • block size step (VLLM_DECODE_BLOCK_BUCKET_STEP): block_size
        • block size max (VLLM_DECODE_BLOCK_BUCKET_MAX): max(128, (max_num_seqs*max_model_len)/block_size)

Additionally, there are HPU PyTorch Bridge environment variables impacting vLLM execution:

  • PT_HPU_LAZY_MODE: if 0, PyTorch Eager backend for Gaudi will be used, if 1 PyTorch Lazy backend for Gaudi will be used, 1 is default
  • PT_HPU_ENABLE_LAZY_COLLECTIVES: required to be true for tensor parallel inference with HPU Graphs

Troubleshooting: Tweaking HPU Graphs

If you experience device out-of-memory issues or want to attempt inference at higher batch sizes, try tweaking HPU Graphs by following the below:

  • Tweak gpu_memory_utilization knob. It will decrease the allocation of KV cache, leaving some headroom for capturing graphs with larger batch size. By default gpu_memory_utilization is set to 0.9. It attempts to allocate ~90% of HBM left for KV cache after short profiling run. Note that decreasing reduces the number of KV cache blocks you have available, and therefore reduces the effective maximum number of tokens you can handle at a given time.
  • If this method is not efficient, you can disable HPUGraph completely. With HPU Graphs disabled, you are trading latency and throughput at lower batches for potentially higher throughput on higher batches. You can do that by adding --enforce-eager flag to server (for online inference), or by passing enforce_eager=True argument to LLM constructor (for offline inference).