[WIP][models][deepseek][qwen2.5] Add support for Qwen2.5 and Deepseek-Distilled-Qwen models

[orionpapadakis]

No description provided.

[Copilot]

Pull Request Overview

This PR adds support for Qwen2.5 and Deepseek-Distilled-Qwen models by implementing the Qwen2 architecture. The implementation reuses the Qwen3 tokenizer and includes special handling for Deepseek-R1-Distill-Qwen models.

• Implements Qwen2 model architecture with configuration, state management, and inference
• Adds model type detection and loading for QWEN_2 and DEEPSEEK_R1_DISTILL_QWEN
• Introduces model-specific behavior flags for system prompts, reasoning tokens, and begin-of-text handling

Reviewed Changes

Copilot reviewed 12 out of 12 changed files in this pull request and generated 5 comments.

Show a summary per file

File	Description
TornadoVMMasterPlan.java	Adds unsupported operation handling for new model types
Qwen3Tokenizer.java	Updates token display logic to include reasoning tokens
Qwen3.java	Adds shouldAddBeginOfText override
Qwen2Configuration.java	Defines configuration structure for Qwen2 models
Qwen2.java	Implements main Qwen2 model class with inference logic
Qwen2ModelLoader.java	Handles loading and weight parsing for Qwen2 models
ModelLoader.java	Adds model type detection for new models
ModelType.java	Defines new model types and loading logic
Model.java	Adds behavior flags and updates interactive/instruction methods
Qwen2StandardWeights.java	Extends StandardWeights with Qwen2-specific bias tensors
Qwen2State.java	Implements state management for Qwen2 models
InferenceCore.java	Adds forwardJavaQwen2 inference implementation

Comments suppressed due to low confidence (1)

src/main/java/com/example/model/Model.java:226

• This line appears to be in the wrong location. It seems to belong in the InferenceCore implementation rather than in the Model interface default method.

                if (tokenizer().shouldDisplayToken(token)) {

[Copilot]

The error message should be more specific about why kvMul() is not supported for Qwen2 models or what alternative should be used.

Suggested change

	throw new UnsupportedOperationException("Not supported for Qwen2.");
	throw new UnsupportedOperationException("kvMul() is not supported for Qwen2 models because they do not use a key-value head multiplier. If you need key-value head information, use numberOfKeyValueHeads instead.");

[Copilot]

The comment contains inconsistent formatting. It should be either '<|beginoftext|>' or '<|begin_of_text|>' to match the actual token format used elsewhere in the codebase.

Suggested change

	* No <|beginoftext|> needed for Qwen models.
	* No <|begin_of_text|> needed for Qwen models.

[Copilot]

The TODO comment indicates incomplete implementation. Either implement the Qwen2-specific fields or remove the comment if no additional fields are needed.

Suggested change

//Qwen2 specific fields TODO

[Copilot]

Another TODO comment indicating incomplete initialization. Either implement the initialization logic or remove the comment if no additional initialization is required.

[Copilot]

The chat tokens configuration uses different Unicode characters for DeepSeek ('｜' vs '|'). This should be documented or extracted to constants to avoid confusion about the character differences.

Suggested change

	new ChatTokens( "<｜begin▁of▁sentence｜>", "", "", "<｜end▁of▁sentence｜>", "") :
	new ChatTokens( "<|im_start|>", "<|im_end|>", "", "<|end_of_text|>", "<|endoftext|>");
	new ChatTokens(DEEPSEEK_BEGIN_OF_SENTENCE, "", "", DEEPSEEK_END_OF_SENTENCE, "") :
	new ChatTokens(QWEN2_IM_START, QWEN2_IM_END, "", QWEN2_END_OF_TEXT, QWEN2_ENDOFTEXT);