[Bug]: guided_json bad output for llama2-13b

[pseudotensor]

Your current environment

Collecting environment information...
PyTorch version: N/A
Is debug build: N/A
CUDA used to build PyTorch: N/A
ROCM used to build PyTorch: N/A

OS: Ubuntu 22.04.3 LTS (x86_64)
GCC version: (Ubuntu 11.4.0-1ubuntu1~22.04) 11.4.0
Clang version: Could not collect
CMake version: Could not collect
Libc version: glibc-2.35

Python version: 3.10.9 (main, Jan 11 2023, 15:21:40) [GCC 11.2.0] (64-bit runtime)
Python platform: Linux-5.15.0-89-generic-x86_64-with-glibc2.35
Is CUDA available: N/A
CUDA runtime version: 12.1.66
CUDA_MODULE_LOADING set to: N/A
GPU models and configuration:
GPU 0: NVIDIA A100-SXM4-80GB
GPU 1: NVIDIA A100-SXM4-80GB
GPU 2: NVIDIA A100-SXM4-80GB
GPU 3: NVIDIA A100-SXM4-80GB
GPU 4: NVIDIA A100-SXM4-80GB
GPU 5: NVIDIA A100-SXM4-80GB
GPU 6: NVIDIA A100-SXM4-80GB
GPU 7: NVIDIA A100-SXM4-80GB

Nvidia driver version: 550.54.14
cuDNN version: Could not collect
HIP runtime version: N/A
MIOpen runtime version: N/A
Is XNNPACK available: N/A

CPU:
Architecture: x86_64
CPU op-mode(s): 32-bit, 64-bit
Address sizes: 46 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
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 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 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 amx_bf16 avx512_fp16 amx_tile amx_int8 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): 0,2,4,6,8,10,12,14,16,18,20,22,24,26,28,30,32,34,36,38,40,42,44,46,48,50,52,54,56,58,60,62,64,66,68,70,72,74,76,78,80,82,84,86,88,90,92,94,96,98,100,102,104,106,108,110,112,114,116,118,120,122,124,126,128,130,132,134,136,138,140,142,144,146,148,150,152,154,156,158,160,162,164,166,168,170,172,174,176,178,180,182,184,186,188,190
NUMA node1 CPU(s): 1,3,5,7,9,11,13,15,17,19,21,23,25,27,29,31,33,35,37,39,41,43,45,47,49,51,53,55,57,59,61,63,65,67,69,71,73,75,77,79,81,83,85,87,89,91,93,95,97,99,101,103,105,107,109,111,113,115,117,119,121,123,125,127,129,131,133,135,137,139,141,143,145,147,149,151,153,155,157,159,161,163,165,167,169,171,173,175,177,179,181,183,185,187,189,191
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] No relevant packages
[conda] No relevant packagesROCM Version: Could not collect
Neuron SDK Version: N/A
vLLM Version: N/A
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 NIC5 NIC6 CPU Affinity NUMA Affinity GPU NUMA ID
GPU0 X NV12 NV12 NV12 NV12 NV12 NV12 NV12 PXB PXB PXB SYS SYS SYS SYS 0,2,4,6,8,10 0 N/A
GPU1 NV12 X NV12 NV12 NV12 NV12 NV12 NV12 SYS SYS SYS SYS SYS SYS SYS 0,2,4,6,8,10 0 N/A
GPU2 NV12 NV12 X NV12 NV12 NV12 NV12 NV12 SYS SYS SYS SYS SYS SYS SYS 0,2,4,6,8,10 0 N/A
GPU3 NV12 NV12 NV12 X NV12 NV12 NV12 NV12 SYS SYS SYS PXB SYS SYS SYS 0,2,4,6,8,10 0 N/A
GPU4 NV12 NV12 NV12 NV12 X NV12 NV12 NV12 SYS SYS SYS SYS PXB SYS PXB 1,3,5,7,9,11 1 N/A
GPU5 NV12 NV12 NV12 NV12 NV12 X NV12 NV12 SYS SYS SYS SYS SYS SYS SYS 1,3,5,7,9,11 1 N/A
GPU6 NV12 NV12 NV12 NV12 NV12 NV12 X NV12 SYS SYS SYS SYS SYS PXB SYS 1,3,5,7,9,11 1 N/A
GPU7 NV12 NV12 NV12 NV12 NV12 NV12 NV12 X SYS SYS SYS SYS SYS SYS SYS 1,3,5,7,9,11 1 N/A
NIC0 PXB SYS SYS SYS SYS SYS SYS SYS X PIX PIX SYS SYS SYS SYS
NIC1 PXB SYS SYS SYS SYS SYS SYS SYS PIX X PIX SYS SYS SYS SYS
NIC2 PXB SYS SYS SYS SYS SYS SYS SYS PIX PIX X SYS SYS SYS SYS
NIC3 SYS SYS SYS PXB SYS SYS SYS SYS SYS SYS SYS X SYS SYS SYS
NIC4 SYS SYS SYS SYS PXB SYS SYS SYS SYS SYS SYS SYS X SYS PIX
NIC5 SYS SYS SYS SYS SYS SYS PXB SYS SYS SYS SYS SYS SYS X SYS
NIC6 SYS SYS SYS SYS PXB SYS SYS SYS SYS SYS SYS SYS PIX 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_2
  NIC3: mlx5_3
  NIC4: mlx5_4
  NIC5: mlx5_7
  NIC6: mlx5_bond_0

🐛 Describe the bug

Using vllm 0.4.0.post1

import openai

client = openai.OpenAI(base_url='')

kwargs = {'model': 'h2oai/h2ogpt-4096-llama2-13b-chat', 'prompt': '[INST] \n\nGive an example employee profile. [/INST]', 'temperature': 0, 'max_tokens': 1024, 'top_p': 1, 'frequency_penalty': 0, 'seed': 1, 'n': 1, 'presence_penalty': 0.14000000000000012, 'stop': ['[/INST]', '', '[INST]'], 'extra_body': {'stop_token_ids': [2, 2], 'response_format': {'type': 'json_object'}, 'guided_json': {'type': 'object', 'properties': {'name': {'type': 'string'}, 'age': {'type': 'integer'}, 'skills': {'type': 'array', 'items': {'type': 'string', 'maxLength': 10}, 'minItems': 3}, 'workhistory': {'type': 'array', 'items': {'type': 'object', 'properties': {'company': {'type': 'string'}, 'duration': {'type': 'string'}, 'position': {'type': 'string'}}, 'required': ['company', 'position']}}}, 'required': ['name', 'age', 'skills', 'workhistory']}}, 'stream': False, 'timeout': 600}
response = client.completions.create(**kwargs).choices[0].text
print(response)

streaming or not doesn't matter, happens every time same way. Get this response:

{

"age":

....

Here ... means a MASSIVE number of new lines.

I've tried about 10 other models, all are ok except llama2-13b.

Others that work:

['h2oai/h2ogpt-4096-llama2-70b-chat',
'mistralai/Mixtral-8x7B-Instruct-v0.1',
                     'HuggingFaceH4/zephyr-7b-beta', 'mistralai/Mistral-7B-Instruct-v0.2', 'openchat/openchat-3.5-1210',
                     'h2oai/h2ogpt-32k-codellama-34b-instruct', 'NousResearch/Nous-Capybara-34B',
                     'h2oai/h2o-danube2-1.8b-chat',
                     'google/gemma-1.1-7b-it',
                     'databricks/dbrx-instruct', 'CohereForAI/c4ai-command-r-v01',]

[simon-mo]

Thank you for the report. I think this is because newline is allowed and the model can indefinitely spam new lines if it choose to be. This commonly occurs when the model is not well prompted or finetuned for this workload. Even with guided JSON, give a few shot examples could be helpful. So prompting would be the simplest fix.

I can think of two slightly more complex fix:

• Some where in outlines (cc @rlouf) whitespace repeated is allowed between JSON key and values, maybe we should just set it to one?
• Maybe we can consider repetition penalty, but this will change output accuracy as it will also apply things like values.

[pseudotensor]

Thanks, understood, but so many other models have no issue with the same prompt. So while prompting might help, would be great to use this as edge case that could make vLLM more reliable in general.

I noticed in the vllm test that you put the json template in the prompt itself. I suppose that might help, but I didn't notice it as required for any other models.

Sounds like a reasonable solution to not allow repeated white space, no usefulness and not an issue with strict json following.

[pseudotensor]

I confirmed that adding to the prompt the prompt template seems to work, but I recommend still having the regexp that leads to guided_json such that multilpe white spaces are not allowed between keys and values and other such non-quoted locations.

[br3no]

You need to instruct the model exactly what you expect it to do. The guided_json parameter (and siblings) will only enforce the generation matches the spec. You still need to nudge the LLM to produce an output in the format you want by adding it in your prompt.

[pseudotensor]

@br3no That's fine, but other models have no issues with following the schema, so I feel like this is an issue that could happen to any model potentially.

There's absolutely no reason for the regexp going into the guided_json to allow arbitrary amounts of white space between key and values.

[rlouf]

There's absolutely no reason for the regexp going into the guided_json to allow arbitrary amounts of white space between key and values.

Absolutely no reason? Unless you know with certainty what models have seen during training the most reasonable thing you can do is assume inputs followed the JSON spec more generally: https://www.json.org/json-en.html

[pseudotensor]

@rlouf Your webpage doesn't say anything relevant to the amount of white space.

My point is that JSON is defined so that the amount of white space does not matter, so might as well restrict generation to only include the minimal amount of white space required to constraint the generation.

[pseudotensor]

E.g.

https://www.ibm.com/docs/SSLTBW_2.4.0/com.ibm.zos.v2r4.ceea900/igz0337w.htm#:~:text=An%20acceptable%20JSON%20white%20space,(Unicode%20U+000D).

i.e. Use one space after the name-separator (colon)

In this case, vLLM did not do this.

[rlouf]

@rlouf Your webpage doesn't say anything relevant to the amount of white space.

It gives the grammar, which defines what is considered valid JSON.

My point is that JSON is defined so that the amount of white space does not matter, so might as well restrict generation to only include the minimal amount of white space required to constraint the generation.

To my point, do you know what models have seen during training? What if you impose this restriction and models have seen something different? What if different model training processes pre-process JSON differently or what if they don't pre-process it at all?

[pseudotensor]

@rlouf I've provided a real case that shows a problem with how it's done right now and there is a solution. Your statement is just speculative questions without any evidence showing any problem.

[rlouf]

Are you suggesting we should only accommodate one particular use case?

[pseudotensor]

I suggest we address real problems with reasonable solutions. The proposal to limit the white space between key and value is perfectly compatible with recommendations of how JSON should be constructed.

If we are worried about being too limiting, then ok, limit to no more than 3 spaces. But in my case it make 1000 new lines :) I don't think that's useful lack of constraint, and I also think new lines as white space between key and value are not normal JSON.

[pseudotensor]

@rlouf BTW, amazing project (outlines).

[rlouf]

I suggest we address real problems with reasonable solutions. The proposal to limit the white space between key and value is perfectly compatible with recommendations of how JSON should be constructed.

We could limit to max 4 spaces and one line break, I doubt models have seen many objects with more white spaces and line breaks. Btw you can already test for this in Outlines directly, we have a whitespace_pattern keyword argument to generate.json.

Is prompting working though?

[pseudotensor]

@rlouf It's pros and cons. The problem with prompting is that a schema may be 2000 tokens, and that wastes alot of tokens. I did do that and it works, but I'd prefer not to have to. It's a big con.

Other models has no issue at all without the schema, e.g.:

vllm_base_models = ['h2oai/h2ogpt-4096-llama2-70b-chat',
                     'HuggingFaceH4/zephyr-7b-beta', 'mistralai/Mistral-7B-Instruct-v0.2', 'openchat/openchat-3.5-1210',
                     'h2oai/h2ogpt-32k-codellama-34b-instruct', 'NousResearch/Nous-Capybara-34B',
                     'mistralai/Mixtral-8x7B-Instruct-v0.1',
                     'h2oai/h2o-danube2-1.8b-chat',
                     'google/gemma-1.1-7b-it', 'h2oai/mixtral-gm-rag-experimental-v2',
                     'databricks/dbrx-instruct', 'CohereForAI/c4ai-command-r-v01']

for the test case I considered.

I don't see really any cons to the white space limiting, esp. if we let it go out to 3-4, and avoid new lines as well as part of white space (only one at end as you said).

[rlouf]

Could you give this a try using generate.json in Outlines and play with whitespace_pattern? We have an integration for vLLM's offline interface.

And let's open a discussion in Outlines to not pollute vLLM's maintainers' notifications :)

[robcaulk]

I ran into this problem with Llama3 8b instruct and solved it by running this PR #4305 and setting guided_whitespace_pattern to " ".

[taoisu]

Same problem with phi-3-mini-instruct as well

[pseudotensor]

@rlouf FYI I h it this even with 'mistralai/Mixtral-8x7B-Instruct-v0.1' even when providing the schema to the model. Putting '[ \t\n]' failed and '\n' failed to work, but ' ' did ok like @robcaulk mentioned above.

[pseudotensor]

https://github.com/vllm-project/vllm/pull/4305/files