A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 16, 03:15 UTC
[Misc] Provide correct Pixtral-HF chat template (#11891)
Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>
Cyrus Leung committed
Jan 10, 2025 at 01:19 UTC
9a228348d2f9a2c85dfc67d6b9fe883bf10a4680
3 files changed
+73
-27
docs/source/models/supported_models.md
+34
-27
@@ -322,7 +322,7 @@ See [this page](#generative-models) for more information on how to use generativ
322
- ✅︎
323
- ✅︎
324
* - `Qwen2ForCausalLM`
325
- - Qwen2
325
+ - QwQ, Qwen2
326
- `Qwen/QwQ-32B-Preview`, `Qwen/Qwen2-7B-Instruct`, `Qwen/Qwen2-7B`, etc.
327
- ✅︎
328
- ✅︎
@@ -436,7 +436,7 @@ loaded. See [relevant issue on HF Transformers](https://github.com/huggingface/t
436
```
437
438
If your model is not in the above list, we will try to automatically convert the model using
439
-{func}`vllm.model_executor.models.adapters.as_embedding_model`. By default, the embeddings
439
+{func}`~vllm.model_executor.models.adapters.as_embedding_model`. By default, the embeddings
440
of the whole prompt are extracted from the normalized hidden state corresponding to the last token.
441
442
#### Reward Modeling (`--task reward`)
@@ -468,7 +468,7 @@ of the whole prompt are extracted from the normalized hidden state corresponding
468
```
469
470
If your model is not in the above list, we will try to automatically convert the model using
471
-{func}`vllm.model_executor.models.adapters.as_reward_model`. By default, we return the hidden states of each token directly.
471
+{func}`~vllm.model_executor.models.adapters.as_reward_model`. By default, we return the hidden states of each token directly.
472
473
```{important}
474
For process-supervised reward models such as `peiyi9979/math-shepherd-mistral-7b-prm`, the pooling config should be set explicitly,
@@ -499,7 +499,7 @@ e.g.: `--override-pooler-config '{"pooling_type": "STEP", "step_tag_id": 123, "r
499
```
500
501
If your model is not in the above list, we will try to automatically convert the model using
502
-{func}`vllm.model_executor.models.adapters.as_classification_model`. By default, the class probabilities are extracted from the softmaxed hidden state corresponding to the last token.
502
+{func}`~vllm.model_executor.models.adapters.as_classification_model`. By default, the class probabilities are extracted from the softmaxed hidden state corresponding to the last token.
503
504
#### Sentence Pair Scoring (`--task score`)
505
@@ -550,6 +550,28 @@ On the other hand, modalities separated by `/` are mutually exclusive.
550
551
See [this page](#multimodal-inputs) on how to pass multi-modal inputs to the model.
552
553
+````{important}
554
+To enable multiple multi-modal items per text prompt, you have to set `limit_mm_per_prompt` (offline inference)
555
+or `--limit-mm-per-prompt` (online inference). For example, to enable passing up to 4 images per text prompt:
556
+
557
+Offline inference:
558
+```python
559
+llm = LLM(
560
+ model="Qwen/Qwen2-VL-7B-Instruct",
561
+ limit_mm_per_prompt={"image": 4},
562
+)
563
+```
564
+
565
+Online inference:
566
+```bash
567
+vllm serve Qwen/Qwen2-VL-7B-Instruct --limit-mm-per-prompt image=4
568
+```
569
+````
570
+
571
+```{note}
572
+vLLM currently only supports adding LoRA to the language backbone of multimodal models.
573
+```
574
+
575
### Generative Models
576
577
See [this page](#generative-models) for more information on how to use generative models.
@@ -689,14 +711,14 @@ See [this page](#generative-models) for more information on how to use generativ
711
* - `Phi3VForCausalLM`
712
- Phi-3-Vision, Phi-3.5-Vision
713
- T + I<sup>E+</sup>
692
- - `microsoft/Phi-3-vision-128k-instruct`, `microsoft/Phi-3.5-vision-instruct` etc.
714
+ - `microsoft/Phi-3-vision-128k-instruct`, `microsoft/Phi-3.5-vision-instruct`, etc.
715
-
716
- ✅︎
717
- ✅︎
718
* - `PixtralForConditionalGeneration`
719
- Pixtral
720
- T + I<sup>+</sup>
699
- - `mistralai/Pixtral-12B-2409`, `mistral-community/pixtral-12b` etc.
721
+ - `mistralai/Pixtral-12B-2409`, `mistral-community/pixtral-12b` (see note), etc.
722
-
723
- ✅︎
724
- ✅︎
@@ -715,7 +737,7 @@ See [this page](#generative-models) for more information on how to use generativ
737
- ✅︎
738
- ✅︎
739
* - `Qwen2VLForConditionalGeneration`
718
- - Qwen2-VL
740
+ - QVQ, Qwen2-VL
741
- T + I<sup>E+</sup> + V<sup>E+</sup>
742
- `Qwen/QVQ-72B-Preview`, `Qwen/Qwen2-VL-7B-Instruct`, `Qwen/Qwen2-VL-72B-Instruct`, etc.
743
- ✅︎
@@ -733,26 +755,6 @@ See [this page](#generative-models) for more information on how to use generativ
755
<sup>E</sup> Pre-computed embeddings can be inputted for this modality.
756
<sup>+</sup> Multiple items can be inputted per text prompt for this modality.
757
736
-````{important}
737
-To enable multiple multi-modal items per text prompt, you have to set `limit_mm_per_prompt` (offline inference)
738
-or `--limit-mm-per-prompt` (online inference). For example, to enable passing up to 4 images per text prompt:
739
-
740
-```python
741
-llm = LLM(
742
- model="Qwen/Qwen2-VL-7B-Instruct",
743
- limit_mm_per_prompt={"image": 4},
744
-)
745
-```
746
-
747
-```bash
748
-vllm serve Qwen/Qwen2-VL-7B-Instruct --limit-mm-per-prompt image=4
749
-```
750
-````
751
-
752
-```{note}
753
-vLLM currently only supports adding LoRA to the language backbone of multimodal models.
754
-```
755
-
758
```{note}
759
To use `TIGER-Lab/Mantis-8B-siglip-llama3`, you have pass `--hf_overrides '{"architectures": ["MantisForConditionalGeneration"]}'` when running vLLM.
760
```
@@ -762,6 +764,11 @@ The official `openbmb/MiniCPM-V-2` doesn't work yet, so we need to use a fork (`
764
For more details, please see: <gh-pr:4087#issuecomment-2250397630>
765
```
766
767
+```{note}
768
+The chat template for Pixtral-HF is incorrect (see [discussion](https://huggingface.co/mistral-community/pixtral-12b/discussions/22)).
769
+A corrected version is available at <gh-file:examples/template_pixtral_hf.jinja>.
770
+```
771
+
772
### Pooling Models
773
774
See [this page](pooling-models) for more information on how to use pooling models.
examples/template_pixtral_hf.jinja
new
+38
@@ -0,0 +1,38 @@
1
+{%- if messages[0]["role"] == "system" %}
2
+ {%- set system_message = messages[0]["content"] %}
3
+ {%- set loop_messages = messages[1:] %}
4
+{%- else %}
5
+ {%- set loop_messages = messages %}
6
+{%- endif %}
7
+
8
+{{- bos_token }}
9
+{%- for message in loop_messages %}
10
+ {%- if (message['role'] == 'user') != (loop.index0 % 2 == 0) %}
11
+ {{- raise_exception('After the optional system message, conversation roles must alternate user/assistant/user/assistant/...') }}
12
+ {%- endif %}
13
+ {%- if message["role"] == "user" %}
14
+ {%- if loop.last and system_message is defined %}
15
+ {{- "[INST]" + system_message + "\n" }}
16
+ {%- else %}
17
+ {{- "[INST]" }}
18
+ {%- endif %}
19
+ {%- if message["content"] is not string %}
20
+ {%- for chunk in message["content"] %}
21
+ {%- if chunk["type"] == "text" %}
22
+ {{- chunk["text"] }}
23
+ {%- elif chunk["type"] == "image" %}
24
+ {{- "[IMG]" }}
25
+ {%- else %}
26
+ {{- raise_exception("Unrecognized content type!") }}
27
+ {%- endif %}
28
+ {%- endfor %}
29
+ {%- else %}
30
+ {{- message["content"] }}
31
+ {%- endif %}
32
+ {{- "[/INST]" }}
33
+ {%- elif message["role"] == "assistant" %}
34
+ {{- message["content"] + eos_token}}
35
+ {%- else %}
36
+ {{- raise_exception("Only user and assistant roles are supported, with the exception of an initial optional system message!") }}
37
+ {%- endif %}
38
+{%- endfor %}
tests/entrypoints/test_chat_utils.py
+1
@@ -758,6 +758,7 @@ def test_resolve_content_format_hf_defined(model, expected_format):
758
("template_falcon.jinja", "string"),
759
("template_inkbot.jinja", "string"),
760
("template_llava.jinja", "string"),
761
+ ("template_pixtral_hf.jinja", "openai"),
762
("template_vlm2vec.jinja", "openai"),
763
("tool_chat_template_granite_20b_fc.jinja", "string"),
764
("tool_chat_template_hermes.jinja", "string"),