| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | # Weight Shapes are in the format |
| 5 | # ([K, N], TP_SPLIT_DIM) |
| 6 | # Example: |
| 7 | # A shape of ([14336, 4096], 0) indicates the following GEMM shape, |
| 8 | # - TP1 : K = 14336, N = 4096 |
| 9 | # - TP2 : K = 7168, N = 4096 |
| 10 | # A shape of ([4096, 6144], 1) indicates the following GEMM shape, |
| 11 | # - TP1 : K = 4096, N = 6144 |
| 12 | # - TP4 : K = 4096, N = 1536 |
| 13 | |
| 14 | # TP1 shapes |
| 15 | WEIGHT_SHAPES = { |
| 16 | "mistralai/Mistral-7B-v0.1": [ |
| 17 | ([4096, 6144], 1), |
| 18 | ([4096, 4096], 0), |
| 19 | ([4096, 28672], 1), |
| 20 | ([14336, 4096], 0), |
| 21 | ], |
| 22 | "meta-llama/Llama-2-7b-hf": [ |
| 23 | ([4096, 12288], 1), |
| 24 | ([4096, 4096], 0), |
| 25 | ([4096, 22016], 1), |
| 26 | ([11008, 4096], 0), |
| 27 | ], |
| 28 | "meta-llama/Llama-3-8b": [ |
| 29 | ([4096, 6144], 1), |
| 30 | ([4096, 4096], 0), |
| 31 | ([4096, 28672], 1), |
| 32 | ([14336, 4096], 0), |
| 33 | ], |
| 34 | "meta-llama/Llama-2-13b-hf": [ |
| 35 | ([5120, 15360], 1), |
| 36 | ([5120, 5120], 0), |
| 37 | ([5120, 27648], 1), |
| 38 | ([13824, 5120], 0), |
| 39 | ], |
| 40 | "meta-llama/Llama-2-70b-hf": [ |
| 41 | ([8192, 10240], 1), |
| 42 | ([8192, 8192], 0), |
| 43 | ([8192, 57344], 1), |
| 44 | ([28672, 8192], 0), |
| 45 | ], |
| 46 | "meta-llama/Llama-3.1-405b-hf": [ |
| 47 | ([16384, 18432], 1), |
| 48 | ([16384, 16384], 0), |
| 49 | ([16384, 106496], 1), |
| 50 | ([53248, 16384], 0), |
| 51 | ], |
| 52 | "meta-llama/Llama-3.1-8B-Instruct": [ |
| 53 | ([4096, 6144], 1), |
| 54 | ([4096, 4096], 0), |
| 55 | ([4096, 28672], 1), |
| 56 | ([14336, 4096], 0), |
| 57 | ], |
| 58 | "meta-llama/Llama-3.3-70B-Instruct": [ |
| 59 | ([8192, 10240], 1), |
| 60 | ([8192, 8192], 0), |
| 61 | ([8192, 57344], 1), |
| 62 | ([28672, 8192], 0), |
| 63 | ], |
| 64 | "mistralai/Mistral-Large-Instruct-2407": [ |
| 65 | ([12288, 14336], 1), |
| 66 | ([12288, 12288], 0), |
| 67 | ([12288, 57344], 1), |
| 68 | ([28672, 12288], 0), |
| 69 | ], |
| 70 | "Qwen/Qwen2.5-7B-Instruct": [ |
| 71 | ([3584, 4608], 1), |
| 72 | ([3584, 3584], 0), |
| 73 | ([3584, 37888], 1), |
| 74 | ([18944, 3584], 0), |
| 75 | ], |
| 76 | "Qwen/Qwen2.5-32B-Instruct": [ |
| 77 | ([5120, 7168], 1), |
| 78 | ([5120, 5120], 0), |
| 79 | ([5120, 55296], 1), |
| 80 | ([27648, 5120], 0), |
| 81 | ], |
| 82 | "Qwen/Qwen2.5-72B-Instruct": [ |
| 83 | ([8192, 10240], 1), |
| 84 | ([8192, 8192], 0), |
| 85 | ([8192, 59136], 1), |
| 86 | ([29568, 8192], 0), |
| 87 | ], |
| 88 | "deepseek-ai/DeepSeek-Coder-V2-Lite-Instruct": [ |
| 89 | ([2048, 3072], 1), |
| 90 | ([2048, 4096], 1), |
| 91 | ([2048, 2048], 0), |
| 92 | ([2048, 576], 0), |
| 93 | ([2048, 21888], 1), |
| 94 | ([10944, 2048], 0), |
| 95 | ([2048, 2816], 1), |
| 96 | ([1408, 2048], 0), |
| 97 | ], |
| 98 | "CohereLabs/c4ai-command-a-03-2025": [ |
| 99 | ([12288, 14336], 1), |
| 100 | ([12288, 12288], 0), |
| 101 | ([12288, 73728], 1), |
| 102 | ([36864, 12288], 0), |
| 103 | ], |
| 104 | } |