search

vllm-project / vllm

A high-throughput and memory-efficient inference and serving engine for LLMs
https://docs.vllm.ai
Apache License 2.0
26.1k stars 3.82k forks source link

[Performance]: Medusa SD have poor performance than baseline #6777

Open cwlseu opened 1 month ago

cwlseu commented 1 month ago

Proposal to improve performance

Test new feature medusa speculative sampling with vllm v0.5.2. After using Medusa speculative sampling, the performance dropped significantly(mistral 7b tokens qps from 133 -> 93). Could it be that the model formats supported by vLLM are incompatible with those trained in FastDecoding? The same model in TensorRT-LLM achieves a 2x speedup.

model download: https://huggingface.co/FasterDecoding/medusa-1.0-zephyr-7b-beta

Report of performance regression

Llama 13 B 84 tokens/s -> 66 tokens/s

Without speculation

Processed prompts: 100%|███████████████████| 1/1 [00:01<00:00, 1.72s/it, est. speed input: 26.21 toks/s, output: 83.88 toks/s] Processed prompts: 100%|███████████████████| 1/1 [00:01<00:00, 1.70s/it, est. speed input: 26.40 toks/s, output: 84.48 toks/s] Processed prompts: 100%|███████████████████| 1/1 [00:01<00:00, 1.71s/it, est. speed input: 26.37 toks/s, output: 84.39 toks/s] 0.011855694982740615 text: 'The current president of the United States is Joe Biden. He was inaugurated on January 20, 2021, and is the 46th president of the United States. Prior to his presidency, Biden served as a United States Senator from Delaware for 36 years and as the 47th Vice President of the United States under President Barack Obama. As president, Biden has focused on addressing the COVID-19 pandemic, promoting racial justice and equity, and addressing climate change. He has also emphasized the importance of bipartisanship and working with Congress to pass legislation that benefits the American people.'

With speculation

Processed prompts: 100%|███████████████████| 1/1 [00:02<00:00, 2.37s/it, est. speed input: 19.02 toks/s, output: 60.85 toks/s] Processed prompts: 100%|███████████████████| 1/1 [00:02<00:00, 2.17s/it, est. speed input: 20.73 toks/s, output: 66.35 toks/s] Processed prompts: 100%|███████████████████| 1/1 [00:02<00:00, 2.17s/it, est. speed input: 20.72 toks/s, output: 66.30 toks/s] 0.01508857806523641 text: 'The current president of the United States is Joe Biden. He was inaugurated on January 20, 2021, and is the 46th president of the United States. Prior to his presidency, Biden served as a United States Senator from Delaware for 36 years and as the 47th Vice President of the United States under President Barack Obama. As president, Biden has focused on addressing the COVID-19 pandemic, promoting racial justice and equity, and addressing climate change. He has also emphasized the importance of bipartisanship and working with Congress to pass legislation that benefits the American people.'

Mistral 7B: 133 tokens/s -> 93 tokens/s

Without speculation

Processed prompts: 100%|██████████████████| 1/1 [00:01<00:00, 1.98s/it, est. speed input: 21.70 toks/s, output: 129.19 toks/s] Processed prompts: 100%|██████████████████| 1/1 [00:01<00:00, 1.92s/it, est. speed input: 22.34 toks/s, output: 133.02 toks/s] Processed prompts: 100%|██████████████████| 1/1 [00:01<00:00, 1.92s/it, est. speed input: 22.36 toks/s, output: 133.10 toks/s] 0.007516195066273212 text: "As the current president of the United States, Joe Biden has been leading the country through unprecedented times. Elected in November 2020, Biden has faced numerous challenges, including a global pandemic, economic uncertainty, and social unrest. Despite these obstacles, Biden has remained steadfast in his commitment to serving the American people and working to address the issues that matter most to them.\n\nBorn in Scranton, Pennsylvania in 1942, Biden has spent his entire career in public service. He served in the Senate for over three decades, representing Delaware, and was elected as vice president under Barack Obama in 2008. Biden's experience and expertise in government have been invaluable as he navigates the complexities of the presidency.\n\nOne of Biden's top priorities has been addressing the COVID-19 pandemic. He has worked closely with health experts and scientists to develop a comprehensive response plan, which includes a nationwide vaccination campaign, increased testing and contact tracing, and targeted mitigation measures. Biden has also called for a renewed focus on public health infrastructure and resources, recognizing the importance of investing in prevention and preparedness.\n\nEconomic recovery has also"

With speculation

Processed prompts: 0%| | 0/1 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]WARNING 07-25 07:13:58 multi_step.py:57] Prompt logprob is not supported by multi step workers. (e.g., speculative decode uses multi step workers). Processed prompts: 100%|███████████████████| 1/1 [00:03<00:00, 3.27s/it, est. speed input: 13.16 toks/s, output: 78.32 toks/s] Processed prompts: 100%|███████████████████| 1/1 [00:02<00:00, 2.74s/it, est. speed input: 15.67 toks/s, output: 93.31 toks/s] Processed prompts: 100%|███████████████████| 1/1 [00:02<00:00, 2.74s/it, est. speed input: 15.67 toks/s, output: 93.30 toks/s] 0.010721690952777863 text: "As the current president of the United States, Joe Biden has been leading the country through unprecedented times. Elected in November 2020, Biden has faced numerous challenges, including a global pandemic, economic uncertainty, and social unrest. Despite these obstacles, Biden has remained steadfast in his commitment to serving the American people and working to address the issues that matter most to them.\n\nBorn in Scranton, Pennsylvania in 1942, Biden has spent his entire career in public service. He served in the Senate for over three decades, representing Delaware, and was elected as vice president under Barack Obama in 2008. Biden's experience and expertise in government have been invaluable as he navigates the complexities of the presidency.\n\nOne of Biden's top priorities has been addressing the COVID-19 pandemic. He has worked closely with health experts and scientists to develop a comprehensive response plan, which includes a nationwide vaccination campaign, increased testing and contact tracing, and targeted mitigation measures. Biden has also called for a renewed focus on public health infrastructure and resources, recognizing the importance of investing in prevention and preparedness.\n\nEconomic recovery has also"

Misc discussion on performance

No response

Your current environment (if you think it is necessary)

Collecting environment information...
PyTorch version: 2.3.1+cu121
Is debug build: False
CUDA used to build PyTorch: 12.1
ROCM used to build PyTorch: N/A

OS: Ubuntu 22.04.4 LTS (x86_64)
GCC version: (Ubuntu 11.4.0-1ubuntu1~22.04) 11.4.0
Clang version: Could not collect
CMake version: version 3.30.0
Libc version: glibc-2.35

Python version: 3.10.12 (main, Nov 20 2023, 15:14:05) [GCC 11.4.0] (64-bit runtime)
Python platform: Linux-5.15.0-125.006-nvidia-x86_64-with-glibc2.35
Is CUDA available: True
CUDA runtime version: Could not collect
CUDA_MODULE_LOADING set to: LAZY
GPU models and configuration:
GPU 0: NVIDIA H100 80GB HBM3
GPU 1: NVIDIA H100 80GB HBM3
GPU 2: NVIDIA H100 80GB HBM3
GPU 3: NVIDIA H100 80GB HBM3
GPU 4: NVIDIA H100 80GB HBM3
GPU 5: NVIDIA H100 80GB HBM3
GPU 6: NVIDIA H100 80GB HBM3
GPU 7: NVIDIA H100 80GB HBM3

Nvidia driver version: 550.90.07
cuDNN version: Could not collect
HIP runtime version: N/A
MIOpen runtime version: N/A
Is XNNPACK available: True

CPU:
Architecture:                       x86_64
CPU op-mode(s):                     32-bit, 64-bit
Address sizes:                      52 bits physical, 57 bits virtual
Byte Order:                         Little Endian
CPU(s):                             192
On-line CPU(s) list:                0-191
Vendor ID:                          GenuineIntel
Model name:                         Intel(R) Xeon(R) Platinum 8468
CPU family:                         6
Model:                              143
Thread(s) per core:                 2
Core(s) per socket:                 48
Socket(s):                          2
Stepping:                           8
CPU max MHz:                        3800.0000
CPU min MHz:                        800.0000
BogoMIPS:                           4200.00
Flags:                              fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush dts acpi mmx fxsr sse sse2 ss ht tm pbe syscall nx pdpe1gb rdtscp lm constant_tsc art arch_perfmon pebs bts rep_good nopl xtopology nonstop_tsc cpuid aperfmperf tsc_known_freq pni pclmulqdq dtes64 monitor ds_cpl vmx smx est tm2 ssse3 sdbg fma cx16 xtpr pdcm pcid dca sse4_1 sse4_2 x2apic movbe popcnt tsc_deadline_timer aes xsave avx f16c rdrand lahf_lm abm 3dnowprefetch cpuid_fault epb cat_l3 cat_l2 cdp_l3 invpcid_single intel_ppin cdp_l2 ssbd mba ibrs ibpb stibp ibrs_enhanced tpr_shadow vnmi flexpriority ept vpid ept_ad fsgsbase tsc_adjust bmi1 avx2 smep bmi2 erms invpcid cqm rdt_a avx512f avx512dq rdseed adx smap avx512ifma clflushopt clwb intel_pt avx512cd sha_ni avx512bw avx512vl xsaveopt xsavec xgetbv1 xsaves cqm_llc cqm_occup_llc cqm_mbm_total cqm_mbm_local split_lock_detect avx_vnni avx512_bf16 wbnoinvd dtherm ida arat pln pts hwp hwp_act_window hwp_epp hwp_pkg_req avx512vbmi umip pku ospke waitpkg avx512_vbmi2 gfni vaes vpclmulqdq avx512_vnni avx512_bitalg tme avx512_vpopcntdq la57 rdpid bus_lock_detect cldemote movdiri movdir64b enqcmd fsrm md_clear serialize tsxldtrk pconfig arch_lbr avx512_fp16 flush_l1d arch_capabilities
Virtualization:                     VT-x
L1d cache:                          4.5 MiB (96 instances)
L1i cache:                          3 MiB (96 instances)
L2 cache:                           192 MiB (96 instances)
L3 cache:                           210 MiB (2 instances)
NUMA node(s):                       2
NUMA node0 CPU(s):                  0-47,96-143
NUMA node1 CPU(s):                  48-95,144-191
Vulnerability Gather data sampling: Not affected
Vulnerability Itlb multihit:        Not affected
Vulnerability L1tf:                 Not affected
Vulnerability Mds:                  Not affected
Vulnerability Meltdown:             Not affected
Vulnerability Mmio stale data:      Not affected
Vulnerability Retbleed:             Not affected
Vulnerability Spec rstack overflow: Not affected
Vulnerability Spec store bypass:    Mitigation; Speculative Store Bypass disabled via prctl and seccomp
Vulnerability Spectre v1:           Mitigation; usercopy/swapgs barriers and __user pointer sanitization
Vulnerability Spectre v2:           Mitigation; Enhanced IBRS, IBPB conditional, RSB filling, PBRSB-eIBRS SW sequence
Vulnerability Srbds:                Not affected
Vulnerability Tsx async abort:      Not affected

Versions of relevant libraries:
[pip3] flashinfer==0.0.9+cu121torch2.3
[pip3] numpy==1.26.4
[pip3] nvidia-nccl-cu12==2.20.5
[pip3] torch==2.3.1
[pip3] torchvision==0.18.1
[pip3] transformers==4.42.4
[pip3] triton==2.3.1
[conda] Could not collect
ROCM Version: Could not collect
Neuron SDK Version: N/A
vLLM Version: 0.5.2
vLLM Build Flags:
CUDA Archs: Not Set; ROCm: Disabled; Neuron: Disabled
GPU Topology:
GPU0    GPU1    GPU2    GPU3    GPU4    GPU5    GPU6    GPU7    NIC0    NIC1    NIC2    NIC3    NIC4    CPU Affinity    NUMA Affinity   GPU NUMA ID
GPU0     X  NV18    NV18    NV18    NV18    NV18    NV18    NV18    PIX NODE    SYS SYS NODE    0-47,96-143 0N/A
GPU1    NV18     X  NV18    NV18    NV18    NV18    NV18    NV18    PXB NODE    SYS SYS NODE    0-47,96-143 0N/A
GPU2    NV18    NV18     X  NV18    NV18    NV18    NV18    NV18    NODE    PXB SYS SYS NODE    0-47,96-143 0N/A
GPU3    NV18    NV18    NV18     X  NV18    NV18    NV18    NV18    NODE    PIX SYS SYS NODE    0-47,96-143 0N/A
GPU4    NV18    NV18    NV18    NV18     X  NV18    NV18    NV18    SYS SYS PXB NODE    SYS 48-95,144-191   1N/A
GPU5    NV18    NV18    NV18    NV18    NV18     X  NV18    NV18    SYS SYS PIX NODE    SYS 48-95,144-191   1N/A
GPU6    NV18    NV18    NV18    NV18    NV18    NV18     X  NV18    SYS SYS NODE    PXB SYS 48-95,144-191   1N/A
GPU7    NV18    NV18    NV18    NV18    NV18    NV18    NV18     X  SYS SYS NODE    PIX SYS 48-95,144-191   1N/A
NIC0    PIX PXB NODE    NODE    SYS SYS SYS SYS  X  NODE    SYS SYS NODE
NIC1    NODE    NODE    PXB PIX SYS SYS SYS SYS NODE     X  SYS SYS NODE
NIC2    SYS SYS SYS SYS PXB PIX NODE    NODE    SYS SYS  X  NODE    SYS
NIC3    SYS SYS SYS SYS NODE    NODE    PXB PIX SYS SYS NODE     X  SYS
NIC4    NODE    NODE    NODE    NODE    SYS SYS SYS SYS NODE    NODE    SYS SYS  X

Legend:

  X    = Self
  SYS  = Connection traversing PCIe as well as the SMP interconnect between NUMA nodes (e.g., QPI/UPI)
  NODE = Connection traversing PCIe as well as the interconnect between PCIe Host Bridges within a NUMA node
  PHB  = Connection traversing PCIe as well as a PCIe Host Bridge (typically the CPU)
  PXB  = Connection traversing multiple PCIe bridges (without traversing the PCIe Host Bridge)
  PIX  = Connection traversing at most a single PCIe bridge
  NV#  = Connection traversing a bonded set of # NVLinks

NIC Legend:

  NIC0: mlx5_0
  NIC1: mlx5_1
  NIC2: mlx5_4
  NIC3: mlx5_5
  NIC4: mlx5_bond_0
abhigoyal1997 commented 1 month ago

It could be because of model compatibility. Currently, vLLM needs Medusa heads to be in a separate checkpoint dir and formatted in a certain way. Can you please try reformatting the Medusa heads using the model class from here? Here's a reference for how the checkpoint should look: https://huggingface.co/abhigoyal/vllm-medusa-vicuna-7b-v1.3/tree/main.

I'll soon release the heads I trained for Mistral and Llama3 models as well, and try to see if this implementation can be modified to support the format for already trained checkpoints on HF.

CSEEduanyu commented 1 month ago

It could be because of model compatibility. Currently, vLLM needs Medusa heads to be in a separate checkpoint dir and formatted in a certain way. Can you please try reformatting the Medusa heads using the model class from here? Here's a reference for how the checkpoint should look: https://huggingface.co/abhigoyal/vllm-medusa-vicuna-7b-v1.3/tree/main.

I'll soon release the heads I trained for Mistral and Llama3 models as well, and try to see if this implementation can be modified to support the format for already trained checkpoints on HF.

is medusa supported topk? whta is default number

boxiaowave commented 1 month ago

It could be because of model compatibility. Currently, vLLM needs Medusa heads to be in a separate checkpoint dir and formatted in a certain way. Can you please try reformatting the Medusa heads using the model class from here? Here's a reference for how the checkpoint should look: https://huggingface.co/abhigoyal/vllm-medusa-vicuna-7b-v1.3/tree/main.

I'll soon release the heads I trained for Mistral and Llama3 models as well, and try to see if this implementation can be modified to support the format for already trained checkpoints on HF.

hello, thanks for your work on medusa implementation. How do you train your medusa head? Is the training code same with the official version from FlashDecoding which is based on axolotl? In different training code, there are some small difference in medusa model.

coderchem commented 1 week ago

Why is vLLM with Medusa 20% slower than TGI?

ZTianle commented 1 week ago

Why is vLLM with Medusa 20% slower than TGI?

Yep, I've observed similar results that align with your findings and those of others.

@cadedaniel I'm particularly curious about the root causes of this performance gap. Could you please provide any advice or insights on how to optimize vLLM with respect to Medusa speculative decoding? Any insights into factors contributing to this performance difference would be valuable.