Support mlperf inference llama3.1 model (#223)

Author: arjunsuresh

.gitignore

@@ -15,6 +15,10 @@ wheels/
 .coverage
 htmlcov
 *tmp/
-*tmp-ck-*/
+tmp-*
 local/cache/
-
+mlc-log.txt
+repos.json
+index_script.json
+index_cache.json
+index_experiment.json

script/get-ml-model-llama2/meta.yaml

@@ -89,6 +89,9 @@ variations:
     default: true
     env:
       MLC_DOWNLOAD_SRC: mlcommons
+    prehook_deps:
+      - tags: get,rclone-config,_mlperf-llama2
+        force_cache: yes
   hf:
     group: download-source
     env:

script/get-ml-model-llama2/run-rclone.sh

@@ -1,5 +1,3 @@
-rclone config create mlc-llama2 drive config_is_local=false scope=drive.readonly root_folder_id=11tBZvvrh0FCm3XuR5E849K42TqftYdUF
-rclone config reconnect mlc-llama2:
 cmd="rclone sync mlc-llama2:${MLC_GIT_CHECKOUT_FOLDER} ${LLAMA2_CHECKPOINT_PATH}/${MLC_GIT_CHECKOUT_FOLDER} -P"
 echo $cmd
 eval $cmd

script/get-ml-model-llama3/customize.py

@@ -20,6 +20,12 @@ def preprocess(i):
         env['MLC_GIT_CHECKOUT_FOLDER'] = os.path.join(
             path, env['MLC_ML_MODEL_NAME'])
 
+    if env['MLC_DOWNLOAD_SRC'] == "mlcommons":
+        i['run_script_input']['script_name'] = 'run-rclone'
+        if env.get('MLC_OUTDIRNAME', '') != '':
+            env['LLAMA3_CHECKPOINT_PATH'] = env['MLC_OUTDIRNAME']
+        else:
+            env['LLAMA3_CHECKPOINT_PATH'] = os.getcwd()
     env['MLC_TMP_REQUIRE_DOWNLOAD'] = 'yes'
 
     return {'return': 0}

script/get-ml-model-llama3/meta.yaml

@@ -12,6 +12,8 @@ prehook_deps:
 - enable_if_env:
     MLC_TMP_REQUIRE_DOWNLOAD:
     - 'yes'
+    MLC_DOWNLOAD_SRC:
+    - huggingface
   env: {}
   extra_cache_tags: llama3,llama-3
   force_env_keys:
@@ -37,21 +39,48 @@ variations:
       MLC_ML_MODEL_PRECISION: fp16
       MLC_ML_MODEL_WEIGHT_DATA_TYPES: fp16
     group: precision
+  405b:
+    group: model-size
+    default: true
+    env:
+      MLC_ML_MODEL_NAME: Llama-3.1-405B-Instruct
+  8b:
+    group: model-size
+    env:
+      MLC_ML_MODEL_NAME: Llama-3.1-8b-Instruct
+  mlc:
+    group: download-src
+    default: true
+    prehook_deps:
+      - tags: get,rclone-config,_mlperf-llama3-1
+        force_cache: true
+    env:
+      MLC_DOWNLOAD_SRC: mlcommons
+  hf:
+    group: download-src
+    default_variations:
+      huggingface-stub: meta-llama/Llama-3.1-405B-Instruct
+    env:
+      MLC_DOWNLOAD_SRC: huggingface
+
   meta-llama/Llama-3.1-405B-Instruct:
+    base:
+      - 405b
     adr:
       hf-zoo:
         tags: _model-stub.meta-llama/Llama-3.1-405B-Instruct
-    default: true
     env:
-      MLC_ML_MODEL_NAME: Llama-3-405b-instruct
+      MLC_ML_MODEL_NAME: Llama-3.1-405B-Instruct
       MLC_MODEL_ZOO_ENV_KEY: LLAMA3
     group: huggingface-stub
   meta-llama/Llama-3.1-8B-Instruct:
+    base:
+      - 8b
     adr:
       hf-zoo:
         tags: _model-stub.meta-llama/Llama-3.1-8B-Instruct
     env:
-      MLC_ML_MODEL_NAME: Llama-3-8b-instruct
+      MLC_ML_MODEL_NAME: Llama-3.1-8b-Instruct
       MLC_MODEL_ZOO_ENV_KEY: LLAMA3
     group: huggingface-stub
   vllm:

script/get-ml-model-llama3/run-rclone.sh

@@ -0,0 +1,4 @@
+cmd="rclone sync mlc-llama3-1:inference/${MLC_ML_MODEL_NAME} ${LLAMA3_CHECKPOINT_PATH}/${MLC_ML_MODEL_NAME} -P"
+echo $cmd
+eval $cmd
+test $? -eq 0 || exit $?

script/get-rclone-config/customize.py

@@ -14,8 +14,14 @@ def preprocess(i):
 
     quiet = (env.get('MLC_QUIET', False) == 'yes')
 
+    run_cmds = []
     if env.get('MLC_RCLONE_CONFIG_CMD', '') != '':
-        env['MLC_RUN_CMD'] = env['MLC_RCLONE_CONFIG_CMD']
+        run_cmds.append(env['MLC_RCLONE_CONFIG_CMD'])
+
+    if env.get('MLC_RCLONE_CONNECT_CMD', '') != '':
+        run_cmds.append(env['MLC_RCLONE_CONNECT_CMD'])
+
+    env['MLC_RUN_CMD'] = ' && '.join(run_cmds)
 
     return {'return': 0}
 

script/get-rclone-config/meta.yaml

@@ -11,3 +11,11 @@ variations:
   mlc-inference:
     env:
       MLC_RCLONE_CONFIG_CMD: 'rclone config create mlc-inference s3 provider=Cloudflare access_key_id=f65ba5eef400db161ea49967de89f47b secret_access_key=fbea333914c292b854f14d3fe232bad6c5407bf0ab1bebf78833c2b359bdfd2b endpoint=https://c2686074cb2caf5cbaf6d134bdba8b47.r2.cloudflarestorage.com'
+  mlperf-llama2:
+    env:
+      MLC_RCLONE_CONFIG_CMD: 'rclone config create mlc-llama2 drive config_is_local=false scope=drive.readonly root_folder_id=11tBZvvrh0FCm3XuR5E849K42TqftYdUF'
+      MLC_RCLONE_CONNECT_CMD: 'rclone config reconnect mlc-llama2:'
+  mlperf-llama3-1:
+    env:
+      MLC_RCLONE_CONFIG_CMD: 'rclone config create mlc-llama3-1 drive config_is_local=false scope=drive.readonly root_folder_id=12K-2yvmr1ZSZ7SLrhidCbWc0BriN98am'
+      MLC_RCLONE_CONNECT_CMD: 'rclone config reconnect mlc-llama3-1:'