已合并
[pytorch][refact]weight path fix #3633
guihaowen666创建于 2025年11月4日
[pytorch][refact]weight path fix #3633
已合并
从已删除 :br_220_weight_fix_3合入到Ascend/MindSpeed-LLM2.2.0
共 6 个文件变更+9-9
| @@ -105,7 +105,7 @@ | |||
| 105 | "swiglu": null, | 105 | "swiglu": null, |
| 106 | "load": "/data/ci/models/deepseek-v2/mg/deepseek2-l4-tp1-pp1-ep8", | 106 | "load": "/data/ci/models/deepseek-v2/mg/deepseek2-l4-tp1-pp1-ep8", |
| 107 | "tokenizer-type": "PretrainedFromHF", | 107 | "tokenizer-type": "PretrainedFromHF", |
| 108 | - "tokenizer-name-or-path": "/data/gemm/", | 108 | + "tokenizer-name-or-path": "/data/ci/models/deepseek2/hf/deepseek2_hf/", |
| 109 | "bf16": null, | 109 | "bf16": null, |
| 110 | "normalization": "RMSNorm", | 110 | "normalization": "RMSNorm", |
| 111 | "untie-embeddings-and-output-weights": null, | 111 | "untie-embeddings-and-output-weights": null, |
| @@ -185,7 +185,7 @@ | |||
| 185 | "swiglu": null, | 185 | "swiglu": null, |
| 186 | "load": "/data/ci/models/deepseek-v2/mg/deepseek2-l4-tp1-pp1-ep8", | 186 | "load": "/data/ci/models/deepseek-v2/mg/deepseek2-l4-tp1-pp1-ep8", |
| 187 | "tokenizer-type": "PretrainedFromHF", | 187 | "tokenizer-type": "PretrainedFromHF", |
| 188 | - "tokenizer-name-or-path": "/data/gemm/", | 188 | + "tokenizer-name-or-path": "/data/ci/models/deepseek2/hf/deepseek2_hf/", |
| 189 | "bf16": null, | 189 | "bf16": null, |
| 190 | "normalization": "RMSNorm", | 190 | "normalization": "RMSNorm", |
| 191 | "untie-embeddings-and-output-weights": null, | 191 | "untie-embeddings-and-output-weights": null, |
| @@ -11,7 +11,7 @@ WORLD_SIZE=$(($NPUS_PER_NODE*$NNODES)) | |||
| 11 | 11 | ||
| 12 | basepath=$(cd `dirname $0`; cd ../../../; pwd) | 12 | basepath=$(cd `dirname $0`; cd ../../../; pwd) |
| 13 | 13 | ||
| 14 | -CKPT_LOAD_DIR="/data/ci/models/deepseek-v2/mg/deepseek-500b-tp1-pp2-ep2-cp2-overlap-base/" | 14 | +CKPT_LOAD_DIR="/data/ci/models/deepseek-500b/mg/deepseek-500b-tp1-pp2-ep2-cp2-overlap-base/" |
| 15 | DATA_PATH="/data/ci/datasets/processed/pretrain_dataset/alpaca_text_document" | 15 | DATA_PATH="/data/ci/datasets/processed/pretrain_dataset/alpaca_text_document" |
| 16 | TOKENIZER_MODEL="/data/ci/models/llama2/hf/llama-2-7b-hf" | 16 | TOKENIZER_MODEL="/data/ci/models/llama2/hf/llama-2-7b-hf" |
| 17 | 17 | ||
| @@ -12,7 +12,7 @@ WORLD_SIZE=$(($NPUS_PER_NODE*$NNODES)) | |||
| 12 | basepath=$(cd `dirname $0`; cd ../../../; pwd) | 12 | basepath=$(cd `dirname $0`; cd ../../../; pwd) |
| 13 | 13 | ||
| 14 | DATA_PATH="/data/ci/datasets/processed/pretrain_dataset/alpaca_text_document" | 14 | DATA_PATH="/data/ci/datasets/processed/pretrain_dataset/alpaca_text_document" |
| 15 | -TOKENIZER_MODEL="/data/gemm/" | 15 | +TOKENIZER_MODEL="/data/ci/models/deepseek2/hf/deepseek2_hf/" |
| 16 | CKPT_LOAD_DIR=" /data/ci/models/deepseek2/mg/l2_gemm_t1p1e8_new/" | 16 | CKPT_LOAD_DIR=" /data/ci/models/deepseek2/mg/l2_gemm_t1p1e8_new/" |
| 17 | 17 | ||
| 18 | TP=1 | 18 | TP=1 |
| @@ -15,8 +15,8 @@ WORLD_SIZE=$(($NPUS_PER_NODE*$NNODES)) | |||
| 15 | basepath=$(cd `dirname $0`; cd ../../../; pwd) | 15 | basepath=$(cd `dirname $0`; cd ../../../; pwd) |
| 16 | 16 | ||
| 17 | DATA_PATH=/data/ci/datasets/processed/alpaca_dsv3_eod/alpaca_text_document | 17 | DATA_PATH=/data/ci/datasets/processed/alpaca_dsv3_eod/alpaca_text_document |
| 18 | -TOKENIZER_PATH=/data/ci/models/deepseek-500b/mg/deepseek-v3-l4-tp2pp2ep4-dualpipev/ | 18 | +TOKENIZER_PATH=/data/ci/models/deepseek3/mg/deepseek-v3-l4-tp2pp2ep4-dualpipev/ |
| 19 | -CKPT_LOAD_DIR=/data/ci/models/deepseek-500b/mg/deepseek-v3-l4-tp2pp2ep4-dualpipev/ | 19 | +CKPT_LOAD_DIR=/data/ci/models/deepseek3/mg/deepseek-v3-l4-tp2pp2ep4-dualpipev/ |
| 20 | 20 | ||
| 21 | TP=2 | 21 | TP=2 |
| 22 | PP=2 | 22 | PP=2 |
| @@ -6,8 +6,8 @@ SHELL_SCRIPTS_DIR="$BASE_DIR/shell_scripts" | |||
| 6 | BASELINE_DIR="$BASE_DIR/baseline_results" | 6 | BASELINE_DIR="$BASE_DIR/baseline_results" |
| 7 | EXEC_PY_DIR=$(dirname "$BASE_DIR") | 7 | EXEC_PY_DIR=$(dirname "$BASE_DIR") |
| 8 | 8 | ||
| 9 | -GENERATE_LOG_DIR=/data/run_logs | 9 | +GENERATE_LOG_DIR=/data/ci/run_logs |
| 10 | -GENERATE_JSON_DIR=/data/run_jsons | 10 | +GENERATE_JSON_DIR=/data/ci/run_jsons |
| 11 | 11 | ||
| 12 | rm -rf $GENERATE_LOG_DIR/* | 12 | rm -rf $GENERATE_LOG_DIR/* |
| 13 | rm -rf $GENERATE_JSON_DIR/* | 13 | rm -rf $GENERATE_JSON_DIR/* |
| @@ -63,7 +63,7 @@ | |||
| 63 | "swiglu": null, | 63 | "swiglu": null, |
| 64 | "load": "/data/ci/models/deepseek2/mg/deepseek2-l2-tp1-pp1-ep8-new", | 64 | "load": "/data/ci/models/deepseek2/mg/deepseek2-l2-tp1-pp1-ep8-new", |
| 65 | "tokenizer-type": "PretrainedFromHF", | 65 | "tokenizer-type": "PretrainedFromHF", |
| 66 | - "tokenizer-name-or-path": "/data/gemm/", | 66 | + "tokenizer-name-or-path": "/data/ci/models/deepseek2/hf/deepseek2_hf/", |
| 67 | "bf16": null, | 67 | "bf16": null, |
| 68 | "normalization": "RMSNorm", | 68 | "normalization": "RMSNorm", |
| 69 | "untie-embeddings-and-output-weights": null, | 69 | "untie-embeddings-and-output-weights": null, |