models.md

Supported Models

Chat/Instruct Models

• LlaMA-like (LlamaForCausalLM):

  • All LlaMA-1 models
  • LlaMA-2: Chat-7B, etc
  • LlaMA-3: Instruct-8B, Instruct-70B, other derivations such as Llama3-8B-Chinese-Chat
  • LlaMA-3.1: Instruct-8B, Instruct-70B
  • LlaMA-3.2: Instruct-1B, Instruct-3B
  • CodeLlaMA: Instruct-7B (-a CodeLlaMA)
  • LLM-Compiler: 7B, 7B-FTD, 13B, 13B-FTD
  • DeepSeek: Chat-7B (-a DeepSeek) , Coder-6.7B (-a DeepSeekCoder), Coder-Instruct-1.3B (-a DeepSeekCoder) 🔥
  • Yi: (-a Yi)
    • v1: Chat-6B, Chat-34B
    • v1.5: Chat-6B, Chat-9B, Chat-34B, Chat-9B-16K, Chat-34B-16K
    • Coder: Chat-1.5B, Chat-9B
  • WizardLM: LM 7B (-a WizardLM), LM 13B (-a WizardLM), Coder Python-7B (-a WizardCoder)
  • TigerBot: Chat-7B, Chat-13B (-a TigerBot)
  • CodeFuse-DeepSeek: 33B (-a CodeFuseDeepSeek)
  • MAP-Neo: Instruct-7B (-a MAP-Neo)
  • Index: Chat-1.9B, Character-1.9B
  • NuminaMath: 7B-TIR
  • SmolLM: (-a SmolLM)
    • v1: Instruct-1.7B
    • v2: Instruct-1.7B
  • Groq: Llama-3-Groq-8B-Tool-Use (-a Llama-3-Groq-8B-Tool-Use)

  For other models that using LlamaForCausalLM architecture, for example, aiXcoder-7B, try -a Yi.

• Baichuan (BaichuanForCausalLM)

  • Chat-7B, Chat-13B

• ChatGLM (ChatGLMModel):

  • ChatGLM: 6B

  • ChatGLM2 family: ChatGLM2 6B, CodeGeeX2 6B, ChatGLM3 6B

    Tip on CodeGeeX2: Code completion only, no context. Use system prompt to specify language, e.g. -s "# language: python".

  • CharacterGLM: 6B (-a CharacterGLM)

    Note: Use additional key-value pair arguments to specify characters, --kv user_name "..." bot_name "..." user_info "..." bot_info "...".

  • GLM-4: Chat-9B-128k, Chat-9B-1M

  • CodeGeeX4: 9B (-a CodeGeeX4)

• InternLM (InternLMForCausalLM, InternLM2ForCausalLM)

  • v1: Chat-7B, Chat-7B v1.1, Chat-20B
  • v2: Chat-1.8B, Chat-7B, Chat-20B, Math-Plus-1.8B, Math-Plus-7B, Math-Plus-20
  • v2.5: Chat-1.8B, Chat-7B, Chat-7B-1M, Chat-20B

• Mistral (MistralForCausalLM, MixtralForCausalLM)

  • Mistral: Instruct-7B-v0.2, Instruct-7B-v0.3

  • OpenChat: 3.5 (-a OpenChat) 🔥

    Tip: Use system prompt to select modes: -s GPT4 (default mode), -s Math (mathematical reasoning mode).

  • Starling: 7B-beta (-a Starling)

    Note: This is based on OpenChat, and is fully compatible with OpenChat GPT4 mode.

  • WizardLM: Math 7B (-a WizardMath)

  • Mixtral: Instruct-8x7B 🔥, Instruct-8x22B

    Three implementations of sliding-window attention (see SlidingWindowAttentionImpl):

    • Full cache: more RAM is needed.
    • Partial cache: less RAM is needed, and faster than ring cache (default).
    • Ring cache (i.e. rolling cache): least RAM, but current implementation is naive (slow). 💣

    Note: precision of these implementations differs, which causes different results.

  • NeuralBeagle14: 7B (-a NeuralBeagle)

  • WizardLM-2: WizardLM-2-8x22B (official link is gone) (-a WizardLM-2-MoE)

    Note: For MixtralForCausalLM models, --experts ... is supported to select a subset of experts when converting. For example, --experts 0,1,2,3 selects the first 4 experts.

  • Codestral: 22B-v0.1

  • Mistral-Nemo: Nemo-Instruct-2407

• Phi (PhiForCausalLM, Phi3ForCausalLM)

  • Phi-2

    Tip: --temp 0 is recommended. Don't forget to try --format qa.

  • Dolphin Phi-2 (-a DolphinPhi2) 🐬

  • Phi-3: Mini-Instruct-4k, Mini-Instruct-128k, Medium-Instruct-4k, Medium-Instruct-128k

  • Phi-3.5: Mini-Instruct, MoE-Instruct

• QWen (QWenLMHeadModel, Qwen2ForCausalLM, Qwen2MoeForCausalLM)

  • v1: Chat-7B, Chat-14B, QAnything-7B
  • v1.5: Chat-0.5B, Chat-1.8B, Chat-4B, Chat-7B, Chat-14B, CodeQwen-Chat-7B (-a CodeQwen)
  • v1.5 MoE: Chat-A2.7B
  • v2: Instruct-0.5B, Instruct-1.5B, Instruct-7B, Instruct-72B
  • v2 MoE: Instruct-57B-A14B (💣 not tested)
  • v2.5: Instruct-0.5B, Instruct-1.5B, Instruct-7B, Instruct-14B, Instruct-32B, Instruct-72B
  • v2.5-Coder: Instruct-1.5B, Instruct-7B
  • v2.5-Math: Instruct-1.5B, Instruct-7B, Instruct-72B
  • Marco-o1 (-a Marco-o1)
  • QwQ-32B-Preview (-a QwQ)

• BlueLM (BlueLMForCausalLM)

  • Chat-7B, Chat-7B 32K

• Orion (OrionForCausalLM)

  • Chat-14B

• MiniCPM (MiniCPMForCausalLM, MiniCPM3ForCausalLM)

  • DPO-2B, SFT-2B, SFT-1B🔥
  • 2B-128k (Note: --temp 0 is recommended.)
  • MoE-8x2B
  • 4B

• Adept Persimmon (PersimmonForCausalLM)

  • Chat-8B

• Gemma (GemmaForCausalLM)

  • v1.0: Instruct-2B, Instruct-7B
  • v1.1: Instruct-2B, Instruct-7B
  • CodeGemma v1.1: Instruct-7B
  • v2: Instruct-2B, Instruct-9B, Instruct-27B

• Cohere (CohereForCausalLM)

  • C4AI Command-R
  • Aya-23-8B, Aya-23-35B (-a Aya-23, fully compatible with Command-R)

• Zhinao (ZhinaoForCausalLM)

  • Chat-7B-4K, Chat-7B-32K, Chat-7B-360K

• DeepSeek (DeepseekV2ForCausalLM)

  • V2-Chat (💣 not tested), V2-Lite-Chat

  • Coder-V2-Instruct (💣 not tested), Coder-V2-Lite-Instruct

  Two optimization modes are defined: speed (default) and memory. See BaseMLAttention.

• XVERSE (XverseForCausalLM)

  • Chat-7B, Chat-13B, Chat-65B

  Note: Tokenizer's behavior is not 100% identical.

• AllenAI (OlmoeForCausalLM)

  • OLMoE: Instruct-7B

• Granite (GraniteForCausalLM, GraniteMoeForCausalLM)

  • v3.0 Instruct-1B-A400M, Instruct-3B-A800M, Instruct-2B, Instruct-8B

Base Models

Please use --format completion for these models.

• LlaMA-like (LlamaForCausalLM):

  • DeepSeek: Coder-Base-1.3B (-a DeepSeekCoder), Coder-Base-6.7B (-a DeepSeekCoder)

• DeepSeek (DeepseekV2ForCausalLM)

  • Coder-V2-Base (💣 not tested), Coder-V2-Lite-Base

• Mistral (MistralForCausalLM, MixtralForCausalLM)

  • Mistral: Base-7B-v0.1, Base-7B-v0.3

• Gemma (GemmaForCausalLM)

  • CodeGemma v1.1: Base-2B, Base-7B

• Grok-1

  • Base

    About Grok-1.

• StarCoder (Starcoder2ForCausalLM)

  • Base-3B, Base-7B, Base-15B

• Stable-LM (StableLMEpochModel)

  • Code-3B

RAG Models

• Text Embedding (XLMRobertaModel)

  • BCE-Embedding

  • BGE-M3 (-a BGE-M3)

    Note: Only dense embedding is implemented.

• QA Ranking (XLMRobertaForSequenceClassification)

  • BCE-ReRanker
  • BGE-ReRanker-M3 (-a BGE-Reranker-M3)

LoRA Models

These LoRA models have been tested:

• Llama-3-Chinese-8B-Instruct

Special Models

• Meta-AI multi-token prediction models checkpoints

  Download at least one multi-token prediction checkpoint (such as 7B_1T_4). Assume it is stored at /path/to/llama-multi-predict/7B_1T_4. Make sure tokenizer.model is downloaded to /path/to/llama-multi-predict.

  To convert it with -a llama-multi-token-prediction-ckpt:

  python convert.py -i /path/to/llama-multi-predict/7B_1T_4 -o llama-multi.bin -a llama-multi-token-prediction-ckpt

  This is a base model, and remember to use --format completion.

  Tip: Use --kv n_future_tokens N to change number of future tokens, N = [1, 4].