exo
exo copied to clipboard
Published
17 hours ago
•
exo-explore
exo-explore
Does Tinygrad now not support the deepseek models?
Hello,
Glad to find this awesome framework 🚀
I'm trying to deploy deepseek models with exo. I checked the model_cards in the source code and It seems that deepseek models does not yet support running on the tinygrad engine:
model_cards = {
### llama
"llama-3.3-70b": {
"layers": 80,
"repo": {
"MLXDynamicShardInferenceEngine": "mlx-community/Llama-3.3-70B-Instruct-4bit",
"TinygradDynamicShardInferenceEngine": "unsloth/Llama-3.3-70B-Instruct",
},
},
"llama-3.2-1b": {
"layers": 16,
"repo": {
"MLXDynamicShardInferenceEngine": "mlx-community/Llama-3.2-1B-Instruct-4bit",
"TinygradDynamicShardInferenceEngine": "unsloth/Llama-3.2-1B-Instruct",
},
},
"llama-3.2-1b-8bit": {
"layers": 16,
"repo": {
"MLXDynamicShardInferenceEngine": "mlx-community/Llama-3.2-1B-Instruct-8bit",
"TinygradDynamicShardInferenceEngine": "unsloth/Llama-3.2-1B-Instruct",
},
},
"llama-3.2-3b": {
"layers": 28,
"repo": {
"MLXDynamicShardInferenceEngine": "mlx-community/Llama-3.2-3B-Instruct-4bit",
"TinygradDynamicShardInferenceEngine": "unsloth/Llama-3.2-3B-Instruct",
},
},
"llama-3.2-3b-8bit": {
"layers": 28,
"repo": {
"MLXDynamicShardInferenceEngine": "mlx-community/Llama-3.2-3B-Instruct-8bit",
"TinygradDynamicShardInferenceEngine": "unsloth/Llama-3.2-3B-Instruct",
},
},
"llama-3.2-3b-bf16": {
"layers": 28,
"repo": {
"MLXDynamicShardInferenceEngine": "mlx-community/Llama-3.2-3B-Instruct",
"TinygradDynamicShardInferenceEngine": "unsloth/Llama-3.2-3B-Instruct",
},
},
"llama-3.1-8b": {
"layers": 32,
"repo": {
"MLXDynamicShardInferenceEngine": "mlx-community/Meta-Llama-3.1-8B-Instruct-4bit",
"TinygradDynamicShardInferenceEngine": "mlabonne/Meta-Llama-3.1-8B-Instruct-abliterated",
},
},
"llama-3.1-70b": {
"layers": 80,
"repo": {
"MLXDynamicShardInferenceEngine": "mlx-community/Meta-Llama-3.1-70B-Instruct-4bit",
"TinygradDynamicShardInferenceEngine": "NousResearch/Meta-Llama-3.1-70B-Instruct",
},
},
"llama-3.1-70b-bf16": {
"layers": 80,
"repo": {
"MLXDynamicShardInferenceEngine": "mlx-community/Meta-Llama-3.1-70B-Instruct-bf16-CORRECTED",
"TinygradDynamicShardInferenceEngine": "NousResearch/Meta-Llama-3.1-70B-Instruct",
},
},
"llama-3-8b": {
"layers": 32,
"repo": {
"MLXDynamicShardInferenceEngine": "mlx-community/Meta-Llama-3-8B-Instruct-4bit",
"TinygradDynamicShardInferenceEngine": "TriAiExperiments/SFR-Iterative-DPO-LLaMA-3-8B-R",
},
},
"llama-3-70b": {
"layers": 80,
"repo": {
"MLXDynamicShardInferenceEngine": "mlx-community/Meta-Llama-3-70B-Instruct-4bit",
"TinygradDynamicShardInferenceEngine": "TriAiExperiments/SFR-Iterative-DPO-LLaMA-3-70B-R",
},
},
"llama-3.1-405b": { "layers": 126, "repo": { "MLXDynamicShardInferenceEngine": "mlx-community/Meta-Llama-3.1-405B-4bit", }, },
"llama-3.1-405b-8bit": { "layers": 126, "repo": { "MLXDynamicShardInferenceEngine": "mlx-community/Meta-Llama-3.1-405B-Instruct-8bit", }, },
### mistral
"mistral-nemo": { "layers": 40, "repo": { "MLXDynamicShardInferenceEngine": "mlx-community/Mistral-Nemo-Instruct-2407-4bit", }, },
"mistral-large": { "layers": 88, "repo": { "MLXDynamicShardInferenceEngine": "mlx-community/Mistral-Large-Instruct-2407-4bit", }, },
### deepseek
"deepseek-coder-v2-lite": { "layers": 27, "repo": { "MLXDynamicShardInferenceEngine": "mlx-community/DeepSeek-Coder-V2-Lite-Instruct-4bit-mlx", }, },
"deepseek-coder-v2.5": { "layers": 60, "repo": { "MLXDynamicShardInferenceEngine": "mlx-community/DeepSeek-V2.5-MLX-AQ4_1_64", }, },
"deepseek-v3": { "layers": 61, "repo": { "MLXDynamicShardInferenceEngine": "mlx-community/DeepSeek-V3-4bit", }, },
"deepseek-v3-3bit": { "layers": 61, "repo": { "MLXDynamicShardInferenceEngine": "mlx-community/DeepSeek-V3-3bit", }, },
"deepseek-r1": { "layers": 61, "repo": { "MLXDynamicShardInferenceEngine": "mlx-community/DeepSeek-R1-4bit", }, },
"deepseek-r1-3bit": { "layers": 61, "repo": { "MLXDynamicShardInferenceEngine": "mlx-community/DeepSeek-R1-3bit", }, },
### deepseek distills
"deepseek-r1-distill-qwen-1.5b": { "layers": 28, "repo": { "MLXDynamicShardInferenceEngine": "mlx-community/deepseek-r1-distill-qwen-1.5b", }, },
"deepseek-r1-distill-qwen-1.5b-3bit": { "layers": 28, "repo": { "MLXDynamicShardInferenceEngine": "mlx-community/DeepSeek-R1-Distill-Qwen-1.5B-3bit", }, },
"deepseek-r1-distill-qwen-1.5b-6bit": { "layers": 28, "repo": { "MLXDynamicShardInferenceEngine": "mlx-community/DeepSeek-R1-Distill-Qwen-1.5B-6bit", }, },
"deepseek-r1-distill-qwen-1.5b-8bit": { "layers": 28, "repo": { "MLXDynamicShardInferenceEngine": "mlx-community/DeepSeek-R1-Distill-Qwen-1.5B-8bit", }, },
"deepseek-r1-distill-qwen-1.5b-bf16": { "layers": 28, "repo": { "MLXDynamicShardInferenceEngine": "mlx-community/DeepSeek-R1-Distill-Qwen-1.5B-bf16", }, },
"deepseek-r1-distill-qwen-7b": { "layers": 28, "repo": { "MLXDynamicShardInferenceEngine": "mlx-community/DeepSeek-R1-Distill-Qwen-7B-4bit", }, },
"deepseek-r1-distill-qwen-7b-3bit": { "layers": 28, "repo": { "MLXDynamicShardInferenceEngine": "mlx-community/DeepSeek-R1-Distill-Qwen-7B-3bit", }, },
"deepseek-r1-distill-qwen-7b-6bit": { "layers": 28, "repo": { "MLXDynamicShardInferenceEngine": "mlx-community/DeepSeek-R1-Distill-Qwen-7B-6bit", }, },
"deepseek-r1-distill-qwen-7b-8bit": { "layers": 28, "repo": { "MLXDynamicShardInferenceEngine": "mlx-community/DeepSeek-R1-Distill-Qwen-7B-8bit", }, },
"deepseek-r1-distill-qwen-7b-bf16": { "layers": 28, "repo": { "MLXDynamicShardInferenceEngine": "mlx-community/DeepSeek-R1-Distill-Qwen-7B-bf16", }, },
"deepseek-r1-distill-qwen-14b": { "layers": 48, "repo": { "MLXDynamicShardInferenceEngine": "mlx-community/DeepSeek-R1-Distill-Qwen-14B-4bit", }, },
"deepseek-r1-distill-qwen-14b-3bit": { "layers": 48, "repo": { "MLXDynamicShardInferenceEngine": "mlx-community/DeepSeek-R1-Distill-Qwen-14B-3bit", }, },
"deepseek-r1-distill-qwen-14b-6bit": { "layers": 48, "repo": { "MLXDynamicShardInferenceEngine": "mlx-community/DeepSeek-R1-Distill-Qwen-14B-6bit", }, },
"deepseek-r1-distill-qwen-14b-8bit": { "layers": 48, "repo": { "MLXDynamicShardInferenceEngine": "mlx-community/DeepSeek-R1-Distill-Qwen-14B-8bit", }, },
"deepseek-r1-distill-qwen-14b-bf16": { "layers": 48, "repo": { "MLXDynamicShardInferenceEngine": "mlx-community/DeepSeek-R1-Distill-Qwen-14B-bf16", }, },
"deepseek-r1-distill-qwen-32b": { "layers": 64, "repo": { "MLXDynamicShardInferenceEngine": "mlx-community/DeepSeek-R1-Distill-Qwen-32B-4bit", }, },
"deepseek-r1-distill-qwen-32b-3bit": { "layers": 64, "repo": { "MLXDynamicShardInferenceEngine": "mlx-community/DeepSeek-R1-Distill-Qwen-32B-3bit", }, },
"deepseek-r1-distill-qwen-32b-6bit": { "layers": 64, "repo": { "MLXDynamicShardInferenceEngine": "mlx-community/DeepSeek-R1-Distill-Qwen-32B-6bit", }, },
"deepseek-r1-distill-qwen-32b-8bit": { "layers": 64, "repo": { "MLXDynamicShardInferenceEngine": "mlx-community/DeepSeek-R1-Distill-Qwen-32B-MLX-8Bit", }, },
"deepseek-r1-distill-qwen-32b-bf16": { "layers": 64, "repo": { "MLXDynamicShardInferenceEngine": "mlx-community/DeepSeek-R1-Distill-Qwen-32B-bf16", }, },
"deepseek-r1-distill-llama-8b": { "layers": 32, "repo": { "MLXDynamicShardInferenceEngine": "mlx-community/DeepSeek-R1-Distill-Llama-8B-4bit", }, },
"deepseek-r1-distill-llama-8b-3bit": { "layers": 32, "repo": { "MLXDynamicShardInferenceEngine": "mlx-community/DeepSeek-R1-Distill-Llama-8B-3bit", }, },
"deepseek-r1-distill-llama-8b-6bit": { "layers": 32, "repo": { "MLXDynamicShardInferenceEngine": "mlx-community/DeepSeek-R1-Distill-Llama-8B-6bit", }, },
"deepseek-r1-distill-llama-8b-8bit": { "layers": 32, "repo": { "MLXDynamicShardInferenceEngine": "mlx-community/DeepSeek-R1-Distill-Llama-8B-8bit", }, },
"deepseek-r1-distill-llama-8b-bf16": { "layers": 32, "repo": { "MLXDynamicShardInferenceEngine": "mlx-community/DeepSeek-R1-Distill-Llama-8B-bf16", }, },
"deepseek-r1-distill-llama-70b": { "layers": 80, "repo": { "MLXDynamicShardInferenceEngine": "mlx-community/DeepSeek-R1-Distill-Llama-70B-4bit", }, },
"deepseek-r1-distill-llama-70b-3bit": { "layers": 80, "repo": { "MLXDynamicShardInferenceEngine": "mlx-community/DeepSeek-R1-Distill-Llama-70B-3bit", }, },
"deepseek-r1-distill-llama-70b-6bit": { "layers": 80, "repo": { "MLXDynamicShardInferenceEngine": "mlx-community/DeepSeek-R1-Distill-Llama-70B-6bit", }, },
"deepseek-r1-distill-llama-70b-8bit": { "layers": 80, "repo": { "MLXDynamicShardInferenceEngine": "mlx-community/DeepSeek-R1-Distill-Llama-70B-8bit", }, },
### llava
"llava-1.5-7b-hf": { "layers": 32, "repo": { "MLXDynamicShardInferenceEngine": "llava-hf/llava-1.5-7b-hf", }, },
### qwen
"qwen-2.5-0.5b": { "layers": 28, "repo": { "MLXDynamicShardInferenceEngine": "mlx-community/Qwen2.5-0.5B-Instruct-4bit", }, },
"qwen-2.5-1.5b": { "layers": 28, "repo": { "MLXDynamicShardInferenceEngine": "mlx-community/Qwen2.5-1.5B-Instruct-4bit", }, },
"qwen-2.5-coder-1.5b": { "layers": 28, "repo": { "MLXDynamicShardInferenceEngine": "mlx-community/Qwen2.5-Coder-1.5B-Instruct-4bit", }, },
"qwen-2.5-3b": { "layers": 36, "repo": { "MLXDynamicShardInferenceEngine": "mlx-community/Qwen2.5-3B-Instruct-4bit", }, },
"qwen-2.5-coder-3b": { "layers": 36, "repo": { "MLXDynamicShardInferenceEngine": "mlx-community/Qwen2.5-Coder-3B-Instruct-4bit", }, },
"qwen-2.5-7b": { "layers": 28, "repo": { "MLXDynamicShardInferenceEngine": "mlx-community/Qwen2.5-7B-Instruct-4bit", }, },
"qwen-2.5-coder-7b": { "layers": 28, "repo": { "MLXDynamicShardInferenceEngine": "mlx-community/Qwen2.5-Coder-7B-Instruct-4bit", }, },
"qwen-2.5-math-7b": { "layers": 28, "repo": { "MLXDynamicShardInferenceEngine": "mlx-community/Qwen2.5-Math-7B-Instruct-4bit", }, },
"qwen-2.5-14b": { "layers": 48, "repo": { "MLXDynamicShardInferenceEngine": "mlx-community/Qwen2.5-14B-Instruct-4bit", }, },
"qwen-2.5-coder-14b": { "layers": 48, "repo": { "MLXDynamicShardInferenceEngine": "mlx-community/Qwen2.5-Coder-14B-Instruct-4bit", }, },
"qwen-2.5-32b": { "layers": 64, "repo": { "MLXDynamicShardInferenceEngine": "mlx-community/Qwen2.5-32B-Instruct-4bit", }, },
"qwen-2.5-coder-32b": { "layers": 64, "repo": { "MLXDynamicShardInferenceEngine": "mlx-community/Qwen2.5-Coder-32B-Instruct-4bit", }, },
"qwen-2.5-72b": { "layers": 80, "repo": { "MLXDynamicShardInferenceEngine": "mlx-community/Qwen2.5-72B-Instruct-4bit", }, },
"qwen-2.5-math-72b": { "layers": 80, "repo": { "MLXDynamicShardInferenceEngine": "mlx-community/Qwen2.5-Math-72B-Instruct-4bit", }, },
### nemotron
"nemotron-70b": { "layers": 80, "repo": { "MLXDynamicShardInferenceEngine": "mlx-community/nvidia_Llama-3.1-Nemotron-70B-Instruct-HF_4bit", }, },
"nemotron-70b-bf16": { "layers": 80, "repo": { "MLXDynamicShardInferenceEngine": "mlx-community/Llama-3.1-Nemotron-70B-Instruct-HF-bf16", }, },
# gemma
"gemma2-9b": { "layers": 42, "repo": { "MLXDynamicShardInferenceEngine": "mlx-community/gemma-2-9b-it-4bit", }, },
"gemma2-27b": { "layers": 46, "repo": { "MLXDynamicShardInferenceEngine": "mlx-community/gemma-2-27b-it-4bit", }, },
# stable diffusion
"stable-diffusion-2-1-base": { "layers": 31, "repo": { "MLXDynamicShardInferenceEngine": "stabilityai/stable-diffusion-2-1-base" } },
# phi
"phi-3.5-mini": { "layers": 32, "repo": { "MLXDynamicShardInferenceEngine": "mlx-community/Phi-3.5-mini-instruct-4bit", }, },
"phi-4": { "layers": 40, "repo": { "MLXDynamicShardInferenceEngine": "mlx-community/phi-4-4bit", }, },
# dummy
"dummy": { "layers": 8, "repo": { "DummyInferenceEngine": "dummy", }, },
}
I can use ollama run to start the DeepSeek model on Linux. Do Exo and Ollama support different model file formats? Sorry, I don't quite understand the difference between them.
Is there anything I can do to advance the support for the DeepSeek model on Tinygrad?
BTW, does the MLX engine currently only support deepseek distillation models?
i thought i seen deepseek v3 in there when i was testng a mac setup, didnt have resources to run it though
BTW, does the MLX engine currently only support deepseek distillation models?
"deepseek-v3": { "layers": 61, "repo": { "MLXDynamicShardInferenceEngine": "mlx-community/DeepSeek-V3-4bit", }, }, "deepseek-v3-3bit": { "layers": 61, "repo": { "MLXDynamicShardInferenceEngine": "mlx-community/DeepSeek-V3-3bit", }, }, "deepseek-r1": { "layers": 61, "repo": { "MLXDynamicShardInferenceEngine": "mlx-community/DeepSeek-R1-4bit", }, }, "deepseek-r1-3bit": { "layers": 61, "repo": { "MLXDynamicShardInferenceEngine": "mlx-community/DeepSeek-R1-3bit", }, },
BTW, does the MLX engine currently only support deepseek distillation models?
"deepseek-v3": { "layers": 61, "repo": { "MLXDynamicShardInferenceEngine": "mlx-community/DeepSeek-V3-4bit", }, }, "deepseek-v3-3bit": { "layers": 61, "repo": { "MLXDynamicShardInferenceEngine": "mlx-community/DeepSeek-V3-3bit", }, }, "deepseek-r1": { "layers": 61, "repo": { "MLXDynamicShardInferenceEngine": "mlx-community/DeepSeek-R1-4bit", }, }, "deepseek-r1-3bit": { "layers": 61, "repo": { "MLXDynamicShardInferenceEngine": "mlx-community/DeepSeek-R1-3bit", }, },
They are all quantized models.
https://x.com/awnihannun/status/1881412271236346233 The model mentioned in this post are DeepSeek R1 671B 3bit quantization. I'll try it.
Same question. Also noticed that exo uses unsloth/* with TinygradDynamicShardInferenceEngine. Maybe tinygrad could support unsloth/DeepSeek-R1-GGUF.
