Skip to content

feat(AIP-3937): Add RTX PRO 6000 Blackwell GPU support #96

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Draft
kbhos-splunk wants to merge 1 commit into
base: main
Choose a base branch
from
+95 −1
Draft

feat(AIP-3937): Add RTX PRO 6000 Blackwell GPU support #96

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
56 changes: 56 additions & 0 deletions config/configs/applications.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -55,6 +55,14 @@ applications:
max_ongoing_requests: 6
ray_actor_options:
num_gpus: 2
RTX_PRO_6000_BLACKWELL:
autoscaling_config:
max_replicas: {{.Replicas.Gemma431bIt}}
min_replicas: {{.Replicas.Gemma431bIt}}
target_ongoing_requests: 4
max_ongoing_requests: 10
ray_actor_options:
num_gpus: 2
options:
autoscaling_config:
max_replicas: {{.Replicas.Gemma431bIt}}
Expand All @@ -78,6 +86,14 @@ applications:
max_num_batched_tokens: 4096
max_num_seqs: 2
tensor_parallel_size: 2
RTX_PRO_6000_BLACKWELL:
engine_args:
dtype: bfloat16
gpu_memory_utilization: 0.85
max_model_len: 240000
max_num_batched_tokens: 4096
max_num_seqs: 1
tensor_parallel_size: 2
model_config:
openai_serving_config:
chat:
Expand Down Expand Up @@ -162,6 +178,9 @@ applications:
L40S:
ray_actor_options:
num_gpus: 1
RTX_PRO_6000_BLACKWELL:
ray_actor_options:
num_gpus: 1
options:
autoscaling_config:
max_replicas: {{.Replicas.GptOss20b}}
Expand All @@ -178,6 +197,10 @@ applications:
engine_args:
gpu_memory_utilization: 0.95
tensor_parallel_size: 1
RTX_PRO_6000_BLACKWELL:
engine_args:
gpu_memory_utilization: 0.50
tensor_parallel_size: 1
model_config:
openai_serving_config:
chat:
Expand Down Expand Up @@ -251,6 +274,9 @@ applications:
L40S:
ray_actor_options:
num_gpus: 0.075
RTX_PRO_6000_BLACKWELL:
ray_actor_options:
num_gpus: 0.031
options:
autoscaling_config:
max_replicas: {{.Replicas.UaeLarge}}
Expand All @@ -266,6 +292,9 @@ applications:
L40S:
engine_args:
gpu_memory_utilization: 0.075
RTX_PRO_6000_BLACKWELL:
engine_args:
gpu_memory_utilization: 0.031
model_config:
engine_args:
gpu_memory_utilization: 0.15
Expand Down Expand Up @@ -317,6 +346,9 @@ applications:
H100:
ray_actor_options:
num_gpus: 0.005
RTX_PRO_6000_BLACKWELL:
ray_actor_options:
num_gpus: 0.004
options:
autoscaling_config:
max_replicas: {{.Replicas.AllMinilmL6V2}}
Expand All @@ -329,6 +361,9 @@ applications:
H100:
engine_args:
gpu_memory_utilization: 0.005
RTX_PRO_6000_BLACKWELL:
engine_args:
gpu_memory_utilization: 0.004
model_config:
engine_args:
gpu_memory_utilization: 0.01
Expand Down Expand Up @@ -500,6 +535,9 @@ applications:
L40S:
ray_actor_options:
num_gpus: 0.05
RTX_PRO_6000_BLACKWELL:
ray_actor_options:
num_gpus: 0.021
options:
autoscaling_config:
max_replicas: {{.Replicas.XlmRobertaLanguageClassifier}}
Expand All @@ -515,6 +553,9 @@ applications:
L40S:
engine_args:
gpu_memory_utilization: 0.05
RTX_PRO_6000_BLACKWELL:
engine_args:
gpu_memory_utilization: 0.021
model_config:
engine_args:
gpu_memory_utilization: 0.1
Expand Down Expand Up @@ -605,6 +646,9 @@ applications:
H100:
ray_actor_options:
num_gpus: 0.005
RTX_PRO_6000_BLACKWELL:
ray_actor_options:
num_gpus: 0.004
options:
autoscaling_config:
max_replicas: {{.Replicas.CrossEncoder}}
Expand All @@ -617,6 +661,9 @@ applications:
H100:
engine_args:
gpu_memory_utilization: 0.005
RTX_PRO_6000_BLACKWELL:
engine_args:
gpu_memory_utilization: 0.004
model_config:
engine_args:
gpu_memory_utilization: 0.01
Expand Down Expand Up @@ -672,6 +719,9 @@ applications:
L40S:
ray_actor_options:
num_gpus: 0.05
RTX_PRO_6000_BLACKWELL:
ray_actor_options:
num_gpus: 0.021
options:
autoscaling_config:
max_replicas: {{.Replicas.E5LanguageClassifier}}
Expand All @@ -687,6 +737,9 @@ applications:
L40S:
engine_args:
gpu_memory_utilization: 0.05
RTX_PRO_6000_BLACKWELL:
engine_args:
gpu_memory_utilization: 0.021
model_config:
engine_args:
gpu_memory_utilization: 0.1
Expand Down Expand Up @@ -741,6 +794,9 @@ applications:
L40S:
ray_actor_options:
num_gpus: 0.025
RTX_PRO_6000_BLACKWELL:
ray_actor_options:
num_gpus: 0.013
options:
autoscaling_config:
max_replicas: {{.Replicas.PromptInjectionCrossEncoder}}
Expand Down
37 changes: 36 additions & 1 deletion config/configs/instance.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -85,4 +85,39 @@ H100_NVL:
cpu: "16"
memory: "48Gi"
ephemeral-storage: "100Gi"
nvidia.com/gpu: "1"
nvidia.com/gpu: "1"
# Keep the key name in sync with applications.yaml gpu_type_options_override keys
# and spec.defaultAcceleratorType — builder.go requires an exact match.
RTX_PRO_6000_BLACKWELL:
- tier: rtx-pro-6000-blackwell-0-gpu
gpusPerPod: 0
env:
NVIDIA_VISIBLE_DEVICES: void
resources:
limits:
cpu: "16"
memory: "24Gi"
ephemeral-storage: "50Gi"
nvidia.com/gpu: "0"
requests:
cpu: "4"
- tier: rtx-pro-6000-blackwell-1-gpu
gpusPerPod: 1
resources:
requests:
cpu: "4"
limits:
cpu: "16"
memory: "48Gi"
ephemeral-storage: "200Gi"
nvidia.com/gpu: "1"
- tier: rtx-pro-6000-blackwell-2-gpu
gpusPerPod: 2
resources:
requests:
cpu: "1"
limits:
cpu: "8"
memory: "96Gi"
ephemeral-storage: "400Gi"
nvidia.com/gpu: "2"
3 changes: 3 additions & 0 deletions tools/cluster_setup/k0s-cluster-config.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -153,6 +153,9 @@ aiPlatform:
name: "splunk-ai-stack"
defaultAcceleratorType: "L40S"
# defaultAcceleratorType: "H100"
# defaultAcceleratorType: "RTX_PRO_6000_BLACKWELL" # RTX PRO 6000 Blackwell (g7e-class) nodes.
# Must exactly match an instance.yaml tier key AND the gpu_type_options_override
# keys in applications.yaml — builder.go errors if no matching worker tier exists.

workerGroupConfig:
imageRegistry: ""
Expand Down
Loading