[Bug]: `"--tokenizer-mode", "mistral"` not compatible with openai API tool use tests

sydnash commented 1 week ago

Your current environment

The output of `python collect_env.py`

```text Collecting environment information... PyTorch version: 2.4.0+cu118 Is debug build: False CUDA used to build PyTorch: 11.8 ROCM used to build PyTorch: N/A OS: Ubuntu 20.04.5 LTS (x86_64) GCC version: (Ubuntu 9.4.0-1ubuntu1~20.04.1) 9.4.0 Clang version: Could not collect CMake version: version 3.22.1 Libc version: glibc-2.31 Python version: 3.10.14 (main, May 6 2024, 19:42:50) [GCC 11.2.0] (64-bit runtime) Python platform: Linux-5.13.0-44-generic-x86_64-with-glibc2.31 Is CUDA available: True CUDA runtime version: 12.4.131 CUDA_MODULE_LOADING set to: LAZY GPU models and configuration: GPU 0: NVIDIA A40 GPU 1: NVIDIA A40 GPU 2: NVIDIA A40 GPU 3: NVIDIA A40 GPU 4: NVIDIA A40 GPU 5: NVIDIA A40 GPU 6: NVIDIA A40 GPU 7: NVIDIA A40 Nvidia driver version: 550.54.15 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 Byte Order: Little Endian Address sizes: 43 bits physical, 48 bits virtual CPU(s): 16 On-line CPU(s) list: 0-15 Thread(s) per core: 1 Core(s) per socket: 16 Socket(s): 1 NUMA node(s): 1 Vendor ID: AuthenticAMD CPU family: 23 Model: 49 Model name: AMD EPYC 7302P 16-Core Processor Stepping: 0 Frequency boost: enabled CPU MHz: 1500.000 CPU max MHz: 3000.0000 CPU min MHz: 1500.0000 BogoMIPS: 6000.00 Virtualization: AMD-V L1d cache: 512 KiB L1i cache: 512 KiB L2 cache: 8 MiB L3 cache: 128 MiB NUMA node0 CPU(s): 0-15 Vulnerability Itlb multihit: Not affected Vulnerability L1tf: Not affected Vulnerability Mds: Not affected Vulnerability Meltdown: 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; Retpolines, IBPB conditional, IBRS_FW, STIBP disabled, RSB filling Vulnerability Srbds: Not affected Vulnerability Tsx async abort: Not affected Flags: fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush mmx fxsr sse sse2 ht syscall nx mmxext fxsr_opt pdpe1gb rdtscp lm constant_tsc rep_good nopl nonstop_tsc cpuid extd_apicid aperfmperf pni pclmulqdq monitor ssse3 fma cx16 sse4_1 sse4_2 movbe popcnt aes xsave avx f16c rdrand lahf_lm cmp_legacy svm extapic cr8_legacy abm sse4a misalignsse 3dnowprefetch osvw ibs skinit wdt tce topoext perfctr_core perfctr_nb bpext perfctr_llc mwaitx cpb cat_l3 cdp_l3 hw_pstate ssbd mba ibrs ibpb stibp vmmcall fsgsbase bmi1 avx2 smep bmi2 cqm rdt_a rdseed adx smap clflushopt clwb sha_ni xsaveopt xsavec xgetbv1 xsaves cqm_llc cqm_occup_llc cqm_mbm_total cqm_mbm_local clzero irperf xsaveerptr rdpru wbnoinvd arat npt lbrv svm_lock nrip_save tsc_scale vmcb_clean flushbyasid decodeassists pausefilter pfthreshold avic v_vmsave_vmload vgif v_spec_ctrl umip rdpid overflow_recov succor smca sme sev sev_es Versions of relevant libraries: [pip3] mypy==1.11.1 [pip3] mypy-extensions==1.0.0 [pip3] numpy==1.26.4 [pip3] nvidia-cublas-cu11==11.11.3.6 [pip3] nvidia-cublas-cu12==12.1.3.1 [pip3] nvidia-cuda-cupti-cu11==11.8.87 [pip3] nvidia-cuda-nvrtc-cu11==11.8.89 [pip3] nvidia-cuda-runtime-cu11==11.8.89 [pip3] nvidia-cudnn-cu11==9.1.0.70 [pip3] nvidia-cudnn-cu12==9.1.0.70 [pip3] nvidia-cufft-cu11==10.9.0.58 [pip3] nvidia-cufft-cu12==11.0.2.54 [pip3] nvidia-curand-cu11==10.3.0.86 [pip3] nvidia-curand-cu12==10.3.2.106 [pip3] nvidia-cusolver-cu11==11.4.1.48 [pip3] nvidia-cusolver-cu12==11.4.5.107 [pip3] nvidia-cusparse-cu11==11.7.5.86 [pip3] nvidia-cusparse-cu12==12.1.0.106 [pip3] nvidia-ml-py==12.560.30 [pip3] nvidia-nccl-cu11==2.20.5 [pip3] nvidia-nccl-cu12==2.20.5 [pip3] nvidia-nvjitlink-cu12==12.6.68 [pip3] nvidia-nvtx-cu11==11.8.86 [pip3] nvidia-nvtx-cu12==12.1.105 [pip3] pyzmq==26.2.0 [pip3] sentence-transformers==3.0.1 [pip3] torch==2.4.0+cu118 [pip3] torchaudio==2.4.0+cu118 [pip3] torchvision==0.19.0+cu118 [pip3] transformers==4.45.1 [pip3] transformers-stream-generator==0.0.5 [pip3] triton==3.0.0 [conda] numpy 1.26.4 pypi_0 pypi [conda] nvidia-cublas-cu11 11.11.3.6 pypi_0 pypi [conda] nvidia-cublas-cu12 12.1.3.1 pypi_0 pypi [conda] nvidia-cuda-cupti-cu11 11.8.87 pypi_0 pypi [conda] nvidia-cuda-nvrtc-cu11 11.8.89 pypi_0 pypi [conda] nvidia-cuda-runtime-cu11 11.8.89 pypi_0 pypi [conda] nvidia-cudnn-cu11 9.1.0.70 pypi_0 pypi [conda] nvidia-cudnn-cu12 9.1.0.70 pypi_0 pypi [conda] nvidia-cufft-cu11 10.9.0.58 pypi_0 pypi [conda] nvidia-cufft-cu12 11.0.2.54 pypi_0 pypi [conda] nvidia-curand-cu11 10.3.0.86 pypi_0 pypi [conda] nvidia-curand-cu12 10.3.2.106 pypi_0 pypi [conda] nvidia-cusolver-cu11 11.4.1.48 pypi_0 pypi [conda] nvidia-cusolver-cu12 11.4.5.107 pypi_0 pypi [conda] nvidia-cusparse-cu11 11.7.5.86 pypi_0 pypi [conda] nvidia-cusparse-cu12 12.1.0.106 pypi_0 pypi [conda] nvidia-ml-py 12.560.30 pypi_0 pypi [conda] nvidia-nccl-cu11 2.20.5 pypi_0 pypi [conda] nvidia-nccl-cu12 2.20.5 pypi_0 pypi [conda] nvidia-nvjitlink-cu12 12.6.68 pypi_0 pypi [conda] nvidia-nvtx-cu11 11.8.86 pypi_0 pypi [conda] nvidia-nvtx-cu12 12.1.105 pypi_0 pypi [conda] pyzmq 26.2.0 pypi_0 pypi [conda] sentence-transformers 3.0.1 pypi_0 pypi [conda] torch 2.4.0+cu118 pypi_0 pypi [conda] torchaudio 2.4.0+cu118 pypi_0 pypi [conda] torchvision 0.19.0+cu118 pypi_0 pypi [conda] transformers 4.45.1 pypi_0 pypi [conda] transformers-stream-generator 0.0.5 pypi_0 pypi [conda] triton 3.0.0 pypi_0 pypi ROCM Version: Could not collect Neuron SDK Version: N/A vLLM Version: 0.1.dev2827+g106909c vLLM Build Flags: CUDA Archs: Not Set; ROCm: Disabled; Neuron: Disabled GPU Topology: GPU0 GPU1 GPU2 GPU3 GPU4 GPU5 GPU6 GPU7 NIC0 NIC1 NIC2 CPU Affinity NUMA Affinity GPU NUMA ID GPU0 X PIX SYS SYS SYS SYS SYS SYS SYS SYS SYS 0-15 0 N/A GPU1 PIX X SYS SYS SYS SYS SYS SYS SYS SYS SYS 0-15 0 N/A GPU2 SYS SYS X PIX SYS SYS SYS SYS PHB SYS SYS 0-15 0 N/A GPU3 SYS SYS PIX X SYS SYS SYS SYS PHB SYS SYS 0-15 0 N/A GPU4 SYS SYS SYS SYS X PIX SYS SYS SYS PHB PHB 0-15 0 N/A GPU5 SYS SYS SYS SYS PIX X SYS SYS SYS PHB PHB 0-15 0 N/A GPU6 SYS SYS SYS SYS SYS SYS X PIX SYS SYS SYS 0-15 0 N/A GPU7 SYS SYS SYS SYS SYS SYS PIX X SYS SYS SYS 0-15 0 N/A NIC0 SYS SYS PHB PHB SYS SYS SYS SYS X SYS SYS NIC1 SYS SYS SYS SYS PHB PHB SYS SYS SYS X PIX NIC2 SYS SYS SYS SYS PHB PHB SYS SYS SYS PIX 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 ```

Model Input Dumps

No response

🐛 Describe the bug

The tests in tests/tool_use for mistral FAILED while add "--tokenizer-mode", "mistral" at the start of vllm serve.

How to reproduce this bug:

Add the CLI to test data in tests/tool_use/utils.py.
Change the tests/tool_use/conftest.py to only test the mistral model.

tests/tool_use/conftest.py:

# for each server config, download the model and return the config
@pytest.fixture(scope="session", params=["mistral"])
def server_config(request):
    config = CONFIGS[request.param]
    # download model and tokenizer using transformers
    snapshot_download(config["model"])
    yield CONFIGS[request.param]

tests/tool_use/utils.py:

 "mistral": {
        "model":
        "mistralai/Mistral-7B-Instruct-v0.3",
        "arguments": [
            "--tool-call-parser", "mistral", "--chat-template",
            str(VLLM_PATH / "examples/tool_chat_template_mistral.jinja"),
            "--ignore-patterns=\"consolidated.safetensors\"",
            "--tokenizer-mode", "mistral"
        ],
        "system_prompt":
        "You are a helpful assistant with access to tools. If a tool"
        " that you have would be helpful to answer a user query, "
        "call the tool. Otherwise, answer the user's query directly "
        "without calling a tool. DO NOT CALL A TOOL THAT IS IRRELEVANT "
        "to the user's question - just respond to it normally."
    }

run the tests in the tests directory:

pytest -v -s tool_use/

The tests will fail.

@patrickvonplaten

I believe this is not just a bug in MistralTokenizer, but rather an incompatibility between the OpenAI API implementation and MistralTokenizer's implementation. Based on my current testing, there are the following four issues:

In MistralTokenizer, there is no vocab attribute, and it should be modified to use the get_vocab method and do not using self.model_tokenizer.tokenizer. to overwrite self.model_tokenizer.


File "/LocalRun/jun.dai/code/github/sydnash/vllm/vllm/entrypoints/openai/tool_parsers/mistral_tool_parser.py", line 48, in __init__
self.bot_token_id = self.model_tokenizer.vocab[self.bot_token]
TypeError: 'method' object is not subscriptable

2. The message parameter of `apply_mistral_chat_template` need to change from `request.messages` to `conversation` due to this error:

ERROR 09-29 17:36:44 serving_chat.py:153] File "/LocalRun/jun.dai/conda/envs/vllm_env/lib/python3.10/site-packages/mistral_common/protocol/instruct/validator.py", line 147, in _validate_assistant_message ERROR 09-29 17:36:44 serving_chat.py:153] raise InvalidAssistantMessageException( ERROR 09-29 17:36:44 serving_chat.py:153] mistral_common.exceptions.InvalidAssistantMessageException: Assistant message must have either content or tool_calls, but not both.

3. The tool call id must change to fit `^[a-zA-Z0-9]{9}$`  regex or modify the tool call id's verify due to this error:

ERROR 09-29 17:29:25 serving_chat.py:153] File "/LocalRun/jun.dai/conda/envs/vllm_env/lib/python3.10/site-packages/mistral_common/protocol/instruct/validator.py", line 310, in _validate_tool_call ERROR 09-29 17:29:25 serving_chat.py:153] raise InvalidFunctionCallException( ERROR 09-29 17:29:25 serving_chat.py:153] mistral_common.exceptions.InvalidFunctionCallException: Tool call id was chatcmpl-tool-03e6481b146e408e9523d9c956696295 but must be a-z, A-Z, 0-9, with a length of 9.

4. The model cannot generate the correct tool call message even doing the above change because the chat_template cannot worked in the `MistralTokenizer`.

[{"name": "get_current_weather", "arguments": {"city": "Dallas", "state": "TX", "unit": "fahrenheit"}}, {"name": "get_current_weather", "arguments": {"city": "Orlando", "state": "FL", "unit": "fahrenheit"}}]


The output message has no `[TOOL_CALLS]`, which is used to identify it as a tool call message.

### Before submitting a new issue...

- [X] Make sure you already searched for relevant issues, and asked the chatbot living at the bottom right corner of the [documentation page](https://docs.vllm.ai/en/latest/), which can answer lots of frequently asked questions.

patrickvonplaten commented 5 days ago

Hey @sydnash,

Did you try to use --tokenizer_format mistral for tool use as shown here: https://github.com/vllm-project/vllm/blob/main/examples/offline_chat_with_tools.py

Mistral's tokenizer cannot work with:

            "--tool-call-parser", "mistral", "--chat-template",
            str(VLLM_PATH / "examples/tool_chat_template_mistral.jinja"),

and that's expected actually

sydnash commented 4 days ago

Hey @sydnash,

Did you try to use --tokenizer_format mistral for tool use as shown here: https://github.com/vllm-project/vllm/blob/main/examples/offline_chat_with_tools.py

Mistral's tokenizer cannot work with:
            "--tool-call-parser", "mistral", "--chat-template",
            str(VLLM_PATH / "examples/tool_chat_template_mistral.jinja"),
and that's expected actually

I modified it like this, but the problem still remains the same.

"arguments": [
            "--tool-call-parser", "mistral", "--chat-template",
            str(VLLM_PATH / "examples/tool_chat_template_mistral.jinja"),
            "--ignore-patterns=\"consolidated.safetensors\"",
            "--tokenizer-mode", "mistral", "--load-format", "mistral", "--config-format", "mistral"
        ],

In the example here https://github.com/vllm-project/vllm/blob/main/examples/offline_chat_with_tools.py you use a random tool call id with nine character, but in the openai API, its tool call id is not a nine character ID consisting only of numbers and letters.

patrickvonplaten commented 4 days ago

you use a random tool call id with nine character

yes that's required when using mistral_common

"arguments": [ "--tool-call-parser", "mistral", "--chat-template", str(VLLM_PATH / "examples/tool_chat_template_mistral.jinja"), "--ignore-patterns=\"consolidated.safetensors\"", "--tokenizer-mode", "mistral", "--load-format", "mistral", "--config-format", "mistral" ],

As said above mistral_common doesn't work with chat templates nor a tool call parser => can you just do:

"arguments": [
            "--ignore-patterns=\"consolidated.safetensors\"",
            "--tokenizer-mode", "mistral", "--load-format", "mistral", "--config-format", "mistral"
        ],

jinja templates are not used in mistral_common

sydnash commented 3 days ago

jinja templates are not used in mistral_common

Yes I know this. This issue is total about that MistralTokenizer tokenizer is not compatible with the openai API implementation in vllm, we can not use this tokenizer while work with the tool call in openai API.

The Mistral model's tool call of OpenAI API currently works without requiring the MistralTokenizer, but I’m unsure if this will continue to be the case because of the warning appears at the start of vllm serve:

FutureWarning: It is strongly recommended to run Mistral models with --tokenizer_mode "mistral" to ensure correct encoding and decoding.

However, using --tokenizer_mode "mistral" will leads to the issues mentioned above, which could be confusing for users. Additionally, we need to consider whether the current OpenAI API implementation should be compatible with MistralTokenizer.

patrickvonplaten commented 1 day ago

Can you give me a command that shows what doesn't work?

sydnash commented 1 day ago