Merge branch 'main' into feature/torchtitan/add-deepseek-model

Xiaoming-AMD · web-flow · commit 3469f9c40314 · 2025-10-20T15:11:58.000+08:00
diff --git a/examples/torchtitan/configs/MI300X/llama3.1_405B-pretrain.yaml b/examples/torchtitan/configs/MI300X/llama3.1_405B-pretrain.yaml
@@ -1,6 +1,6 @@
 work_group: ${PRIMUS_TEAM:amd}
 user_name: ${PRIMUS_USER:root}
-exp_name: ${PRIMUS_EXP_NAME:llama3_405B-pretrain}
+exp_name: ${PRIMUS_EXP_NAME:llama3.1_405B-pretrain}
 workspace: ./output
 
 modules:
diff --git a/examples/torchtitan/configs/MI300X/llama3.1_70B-BF16-pretrain.yaml b/examples/torchtitan/configs/MI300X/llama3.1_70B-BF16-pretrain.yaml
@@ -1,6 +1,6 @@
 work_group: ${PRIMUS_TEAM:amd}
 user_name: ${PRIMUS_USER:root}
-exp_name: ${PRIMUS_EXP_NAME:llama3_70B-pretrain}
+exp_name: ${PRIMUS_EXP_NAME:llama3.1_70B-pretrain}
 workspace: ./output
 
 modules:
diff --git a/examples/torchtitan/configs/MI300X/llama3.1_70B-FP8-pretrain.yaml b/examples/torchtitan/configs/MI300X/llama3.1_70B-FP8-pretrain.yaml
@@ -1,6 +1,6 @@
 work_group: ${PRIMUS_TEAM:amd}
 user_name: ${PRIMUS_USER:root}
-exp_name: ${PRIMUS_EXP_NAME:llama3_70B-pretrain}
+exp_name: ${PRIMUS_EXP_NAME:llama3.1_70B-pretrain}
 workspace: ./output
 
 modules:
diff --git a/examples/torchtitan/configs/MI300X/llama3.1_8B-BF16-pretrain.yaml b/examples/torchtitan/configs/MI300X/llama3.1_8B-BF16-pretrain.yaml
@@ -1,6 +1,6 @@
 work_group: ${PRIMUS_TEAM:amd}
 user_name: ${PRIMUS_USER:root}
-exp_name: ${PRIMUS_EXP_NAME:llama3_8B-pretrain}
+exp_name: ${PRIMUS_EXP_NAME:llama3.1_8B-pretrain}
 workspace: ./output
 
 modules:
diff --git a/examples/torchtitan/configs/MI300X/llama3.1_8B-FP8-pretrain.yaml b/examples/torchtitan/configs/MI300X/llama3.1_8B-FP8-pretrain.yaml
@@ -1,6 +1,6 @@
 work_group: ${PRIMUS_TEAM:amd}
 user_name: ${PRIMUS_USER:root}
-exp_name: ${PRIMUS_EXP_NAME:llama3_8B-pretrain}
+exp_name: ${PRIMUS_EXP_NAME:llama3.1_8B-pretrain}
 workspace: ./output
 
 modules:
diff --git a/primus/configs/models/megatron/llama3_70B.yaml b/primus/configs/models/megatron/llama3_70B.yaml
@@ -2,7 +2,7 @@ bases:
   - llama3_base.yaml
 
 tokenizer_type: Llama3Tokenizer
-tokenizer_model: meta-llama/Llama-3.1-70B
+tokenizer_model: meta-llama/Meta-Llama-3-70B
 
 ffn_hidden_size: 28672
 hidden_size: 8192
diff --git a/primus/configs/models/megatron/llama3_8B.yaml b/primus/configs/models/megatron/llama3_8B.yaml
@@ -2,7 +2,7 @@ bases:
   - llama3_base.yaml
 
 tokenizer_type: Llama3Tokenizer
-tokenizer_model: meta-llama/Llama-3.1-8B
+tokenizer_model: meta-llama/Meta-Llama-3-8B
 
 
 ffn_hidden_size: 14336
diff --git a/primus/configs/models/torchtitan/llama3.1_405B.yaml b/primus/configs/models/torchtitan/llama3.1_405B.yaml
@@ -5,5 +5,4 @@ job:
 model:
   name: "llama3"
   flavor: "405B"
-  hf_assets_path: "meta-llama/Llama-3.1-8B"
-  # converters = ["float8"]
+  hf_assets_path: "meta-llama/Llama-3.1-405B"
diff --git a/primus/configs/models/torchtitan/llama3.1_70B-fp8.yaml b/primus/configs/models/torchtitan/llama3.1_70B-fp8.yaml
@@ -1,6 +1,6 @@
 job:
   dump_folder: "./outputs"
-  description: "Llama 3 70B training"
+  description: "Llama 3.1 70B training"
 
 model:
   name: "llama3"
diff --git a/primus/configs/models/torchtitan/llama3.1_70B.yaml b/primus/configs/models/torchtitan/llama3.1_70B.yaml
@@ -1,9 +1,8 @@
 job:
   dump_folder: "./outputs"
-  description: "Llama 3 70B training"
+  description: "Llama 3.1 70B training"
 
 model:
   name: "llama3"
   flavor: "70B"
-  hf_assets_path: "meta-llama/Llama-3.1-8B"
-  # converters = ["float8"]
+  hf_assets_path: "meta-llama/Llama-3.1-70B"
diff --git a/primus/configs/models/torchtitan/llama3.1_8B-fp8.yaml b/primus/configs/models/torchtitan/llama3.1_8B-fp8.yaml
@@ -1,6 +1,6 @@
 job:
   dump_folder: "./outputs"
-  description: "Llama 3 8B training"
+  description: "Llama 3.1 8B training"
 
 model:
   name: "llama3"
diff --git a/primus/configs/models/torchtitan/llama3.1_8B.yaml b/primus/configs/models/torchtitan/llama3.1_8B.yaml
@@ -1,9 +1,8 @@
 job:
   dump_folder: "./outputs"
-  description: "Llama 3 8B training"
+  description: "Llama 3.1 8B training"
 
 model:
   name: "llama3"
   flavor: "8B"
   hf_assets_path: "meta-llama/Llama-3.1-8B"
-  converters: []
diff --git a/primus/configs/models/torchtitan/llama3.3_70B-fp8.yaml b/primus/configs/models/torchtitan/llama3.3_70B-fp8.yaml
@@ -0,0 +1,11 @@
+job:
+  dump_folder: "./outputs"
+  description: "Llama 3 70B training"
+
+model:
+  name: "llama3"
+  flavor: "70B"
+  hf_assets_path: "meta-llama/Llama-3.3-70B-Instruct"
+  converters:
+    - quantize.linear.float8
+    - quantize.grouped_mm.float8
diff --git a/primus/configs/models/torchtitan/llama3.3_70B.yaml b/primus/configs/models/torchtitan/llama3.3_70B.yaml
@@ -0,0 +1,8 @@
+job:
+  dump_folder: "./outputs"
+  description: "Llama 3 70B training"
+
+model:
+  name: "llama3"
+  flavor: "70B"
+  hf_assets_path: "meta-llama/Llama-3.3-70B-Instruct"
diff --git a/primus/configs/models/torchtitan/llama3_70B-fp8.yaml b/primus/configs/models/torchtitan/llama3_70B-fp8.yaml
@@ -0,0 +1,11 @@
+job:
+  dump_folder: "./outputs"
+  description: "Llama 3 70B training"
+
+model:
+  name: "llama3"
+  flavor: "70B"
+  hf_assets_path: "meta-llama/Meta-Llama-3-70B"
+  converters:
+    - quantize.linear.float8
+    - quantize.grouped_mm.float8
diff --git a/primus/configs/models/torchtitan/llama3_70B.yaml b/primus/configs/models/torchtitan/llama3_70B.yaml
@@ -0,0 +1,8 @@
+job:
+  dump_folder: "./outputs"
+  description: "Llama 3 70B training"
+
+model:
+  name: "llama3"
+  flavor: "70B"
+  hf_assets_path: "meta-llama/Meta-Llama-3-70B"
diff --git a/primus/configs/models/torchtitan/llama3_8B-fp8.yaml b/primus/configs/models/torchtitan/llama3_8B-fp8.yaml
@@ -0,0 +1,11 @@
+job:
+  dump_folder: "./outputs"
+  description: "Llama 3 8B training"
+
+model:
+  name: "llama3"
+  flavor: "8B"
+  hf_assets_path: "meta-llama/Meta-Llama-3-8B"
+  converters:
+    - quantize.linear.float8
+    - quantize.grouped_mm.float8
diff --git a/primus/configs/models/torchtitan/llama3_8B.yaml b/primus/configs/models/torchtitan/llama3_8B.yaml
@@ -0,0 +1,8 @@
+job:
+  dump_folder: "./outputs"
+  description: "Llama 3 8B training"
+
+model:
+  name: "llama3"
+  flavor: "8B"
+  hf_assets_path: "meta-llama/Meta-Llama-3-8B"
