Add Qwen2.5 Instruct Turbo models on Together AI (#3063)

stanford-crfm · Oct 15, 2024 · 1401d10 · 1401d10
1 parent c09fe75
commit 1401d10
Show file tree

Hide file tree

Showing 3 changed files with 38 additions and 0 deletions.
diff --git a/src/helm/config/model_deployments.yaml b/src/helm/config/model_deployments.yaml
@@ -2543,6 +2543,20 @@ model_deployments:
     client_spec:
       class_name: "helm.clients.together_client.TogetherChatClient"
 
+  - name: together/qwen2.5-7b-instruct-turbo
+    model_name: qwen/qwen2.5-7b-instruct-turbo
+    tokenizer_name: qwen/qwen2.5-7b-instruct
+    max_sequence_length: 128000
+    client_spec:
+      class_name: "helm.clients.together_client.TogetherChatClient"
+
+  - name: together/qwen2.5-72b-instruct-turbo
+    model_name: qwen/qwen2.5-72b-instruct-turbo
+    tokenizer_name: qwen/qwen2.5-7b-instruct
+    max_sequence_length: 128000
+    client_spec:
+      class_name: "helm.clients.together_client.TogetherChatClient"
+
   - name: huggingface/qwen-vl
     model_name: qwen/qwen-vl
     tokenizer_name: qwen/qwen-vl

diff --git a/src/helm/config/model_metadata.yaml b/src/helm/config/model_metadata.yaml
@@ -2620,6 +2620,22 @@ models:
     release_date: 2024-06-07
     tags: [TEXT_MODEL_TAG, LIMITED_FUNCTIONALITY_TEXT_MODEL_TAG, INSTRUCTION_FOLLOWING_MODEL_TAG]
 
+  - name: qwen/qwen2.5-7b-instruct-turbo
+    display_name: Qwen2.5 Instruct Turbo (7B)
+    description: Qwen2.5 Instruct Turbo (7B) was trained on 18 trillion tokens and supports 29 languages, and shows improvements over Qwen2 in knowledge, coding, mathematics, instruction following, generating long texts, and processing structure data. ([blog](https://qwenlm.github.io/blog/qwen2.5/)) Turbo is Together's cost-efficient implementation, providing fast FP8 performance while maintaining quality, closely matching FP16 reference models. ([blog](https://www.together.ai/blog/together-inference-engine-2))
+    creator_organization_name: Qwen
+    access: open
+    release_date: 2024-09-19
+    tags: [TEXT_MODEL_TAG, LIMITED_FUNCTIONALITY_TEXT_MODEL_TAG, INSTRUCTION_FOLLOWING_MODEL_TAG]
+
+  - name: qwen/qwen2.5-72b-instruct-turbo
+    display_name: Qwen2.5 Instruct Turbo (72B)
+    description: Qwen2.5 Instruct Turbo (72B) was trained on 18 trillion tokens and supports 29 languages, and shows improvements over Qwen2 in knowledge, coding, mathematics, instruction following, generating long texts, and processing structure data. ([blog](https://qwenlm.github.io/blog/qwen2.5/)) Turbo is Together's cost-efficient implementation, providing fast FP8 performance while maintaining quality, closely matching FP16 reference models. ([blog](https://www.together.ai/blog/together-inference-engine-2))
+    creator_organization_name: Qwen
+    access: open
+    release_date: 2024-09-19
+    tags: [TEXT_MODEL_TAG, LIMITED_FUNCTIONALITY_TEXT_MODEL_TAG, INSTRUCTION_FOLLOWING_MODEL_TAG]
+
   - name: qwen/qwen-vl
     display_name: Qwen-VL
     description: Visual multimodal version of the Qwen large language model series ([paper](https://arxiv.org/abs/2308.12966)).

diff --git a/src/helm/config/tokenizer_configs.yaml b/src/helm/config/tokenizer_configs.yaml
@@ -513,6 +513,14 @@ tokenizer_configs:
     end_of_text_token: <|im_end|>"
     prefix_token: "<|im_start|>'"
 
+  - name: qwen/qwen2.5-7b-instruct
+    tokenizer_spec:
+      class_name: "helm.tokenizers.huggingface_tokenizer.HuggingFaceTokenizer"
+      args:
+        pretrained_model_name_or_path: Qwen/Qwen2.5-7B-Instruct
+    end_of_text_token: <|im_end|>"
+    prefix_token: "<|im_start|>'"
+
   - name: qwen/qwen-vl
     tokenizer_spec:
       class_name: "helm.tokenizers.huggingface_tokenizer.HuggingFaceTokenizer"