Merge remote-tracking branch 'upstream/main'

red-hat-data-services · Feb 28, 2025 · 62c8665 · 62c8665
2 parents c132a8a + bac8162
commit 62c8665
Show file tree

Hide file tree

Showing 7 changed files with 464 additions and 216 deletions.
diff --git a/images/runtime/training/cuda/Pipfile b/images/runtime/training/cuda/Pipfile
@@ -13,7 +13,7 @@ torch = "==2.4.1"
 sentencepiece = "<0.3,>=0.1.99"
 tokenizers = "<1.0,>=0.13.3"
 tqdm = "<5.0,>=4.66.2"
-trl = ">=0.15.1"
+trl = ">=0.15.2"
 protobuf = "<6.0.0,>=5.28.0"
 simpleeval = "<1.0,>=0.9.13"
 safetensors = "*"
@@ -27,6 +27,7 @@ pydantic = ">=2.7.0"
 deepspeed = ">=0.14.3"
 aiofiles = ">=23.2.1"
 async-timeout = "==4.0.3"
+tensorboard = "2.19.0"
 
 [dev-packages]
 

diff --git a/images/runtime/training/cuda/Pipfile.lock b/images/runtime/training/cuda/Pipfile.lock
diff --git a/images/runtime/training/rocm/Dockerfile b/images/runtime/training/rocm/Dockerfile
@@ -64,6 +64,19 @@ RUN micropipenv install && \
     chmod -R g+w /opt/app-root/lib/python3.11/site-packages && \
     fix-permissions /opt/app-root -P
 
+# Install Flash Attention
+ENV GPU_ARCHS=gfx90a;gfx941;gfx942
+
+RUN pip install wheel ninja
+
+RUN export TMP_DIR=$(mktemp -d) \
+    && cd $TMP_DIR \
+    && git clone --depth 1 --branch v2.7.4 https://github.com/Dao-AILab/flash-attention.git \
+    && cd flash-attention \
+    && git submodule update --init
+    && MAX_JOBS="16" python3 setup.py install --verbose \
+    && rm -rf $TMP_DIR
+
 # Restore user workspace
 USER 1001
 

diff --git a/images/runtime/training/rocm/Pipfile b/images/runtime/training/rocm/Pipfile
@@ -19,7 +19,7 @@ pytorch-triton-rocm = {version = "*", index = "pytorch"}
 sentencepiece = "<0.3,>=0.1.99"
 tokenizers = "<1.0,>=0.13.3"
 tqdm = "<5.0,>=4.66.2"
-trl = ">=0.15.1"
+trl = ">=0.15.2"
 protobuf = "<6.0.0,>=5.28.0"
 simpleeval = "<1.0,>=0.9.13"
 safetensors = "*"
@@ -33,6 +33,7 @@ pydantic = ">=2.7.0"
 deepspeed = ">=0.14.3"
 aiofiles = ">=23.2.1"
 async-timeout = "==4.0.3"
+tensorboard = "2.19.0"
 
 [dev-packages]
 

diff --git a/images/runtime/training/rocm/Pipfile.lock b/images/runtime/training/rocm/Pipfile.lock
diff --git a/tests/kfto/kfto_mnist_sdk_test.go b/tests/kfto/kfto_mnist_sdk_test.go
@@ -19,6 +19,7 @@ package kfto
 import (
 	"strings"
 	"testing"
+	"time"
 
 	. "github.com/onsi/gomega"
 	. "github.com/project-codeflare/codeflare-common/support"
@@ -41,10 +42,11 @@ func TestMnistSDK(t *testing.T) {
 	CreateUserRoleBindingWithClusterRole(test, userName, namespace.Name, "admin")
 
 	requiredChangesInNotebook := map[string]string{
-		"${api_url}":   GetOpenShiftApiUrl(test),
-		"${password}":  userToken,
-		"${num_gpus}":  "0",
-		"${namespace}": namespace.Name,
+		"${api_url}":        GetOpenShiftApiUrl(test),
+		"${password}":       userToken,
+		"${num_gpus}":       "0",
+		"${namespace}":      namespace.Name,
+		"${training_image}": GetCudaTrainingImage(),
 	}
 
 	jupyterNotebook := string(readFile(test, "resources/mnist_kfto.ipynb"))
@@ -81,7 +83,7 @@ func TestMnistSDK(t *testing.T) {
 		Should(WithTransform(PyTorchJobConditionRunning, Equal(v1.ConditionTrue)))
 
 	// Make sure that the job eventually succeeds
-	test.Eventually(PyTorchJob(test, namespace.Name, "pytorch-ddp")).
+	test.Eventually(PyTorchJob(test, namespace.Name, "pytorch-ddp"), TestTimeoutLong, 1*time.Second).
 		Should(WithTransform(PyTorchJobConditionSucceeded, Equal(v1.ConditionTrue)))
 
 	// TODO: write torch job logs?

diff --git a/tests/kfto/resources/mnist_kfto.ipynb b/tests/kfto/resources/mnist_kfto.ipynb
@@ -1,5 +1,15 @@
 {
  "cells": [
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "ebdb3af3",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "%pip install -U kubeflow-training"
+   ]
+  },
   {
    "cell_type": "code",
    "execution_count": 6,
@@ -19,7 +29,7 @@
   },
   {
    "cell_type": "code",
-   "execution_count": 7,
+   "execution_count": null,
    "id": "72dd1751",
    "metadata": {},
    "outputs": [],
@@ -28,7 +38,8 @@
     "num_gpus = \"${num_gpus}\"\n",
     "openshift_api_url = \"${api_url}\"\n",
     "namespace = \"${namespace}\"\n",
-    "token = \"${password}\""
+    "token = \"${password}\"\n",
+    "training_image= \"${training_image}\""
    ]
   },
   {
@@ -57,9 +68,9 @@
     "   train_func=train_func,\n",
     "   num_workers=1,\n",
     "   resources_per_worker={\"gpu\": num_gpus},\n",
-    "   base_image=\"quay.io/kpostlet/torch-train:with-minivision\",\n",
-    "   # packages_to_install=[\"torchvision==0.19.0\", \"--target=/tmp/lib\"],\n",
-    "   # env_vars={\"PYTHONPATH\": \"/tmp/lib:$PYTHONPATH\", \"NCCL_DEBUG\": \"INFO\", \"TORCH_DISTRIBUTED_DEBUG\": \"DETAIL\"}\n",
+    "   base_image=training_image,\n",
+    "   packages_to_install=[\"torchvision==0.19.0\",\"minio==7.2.13\", \"--target=/tmp/lib\"],\n",
+    "   env_vars={\"PYTHONPATH\": \"/tmp/lib:$PYTHONPATH\", \"NCCL_DEBUG\": \"INFO\", \"TORCH_DISTRIBUTED_DEBUG\": \"DETAIL\"}\n",
     ")"
    ]
   },