Fixed eGPU deploy?
This commit is contained in:
@@ -23,27 +23,30 @@ job "ai-backend" {
|
|||||||
OLLAMA_HOST = "0.0.0.0"
|
OLLAMA_HOST = "0.0.0.0"
|
||||||
OLLAMA_ORIGINS = "*"
|
OLLAMA_ORIGINS = "*"
|
||||||
|
|
||||||
# CRITICAL FOR 6900XT:
|
# Unlock the 6900XT (Navi 21) for ROCm
|
||||||
# This tells ROCm to treat the card like a supported Pro workstation card
|
|
||||||
HSA_OVERRIDE_GFX_VERSION = "10.3.0"
|
HSA_OVERRIDE_GFX_VERSION = "10.3.0"
|
||||||
|
|
||||||
|
# Debugging enabled so we can confirm it worked
|
||||||
|
OLLAMA_DEBUG = "1"
|
||||||
}
|
}
|
||||||
|
|
||||||
config {
|
config {
|
||||||
# Standard image (contains ROCm libraries)
|
|
||||||
image = "docker.io/ollama/ollama:latest"
|
image = "docker.io/ollama/ollama:latest"
|
||||||
ports = ["api"]
|
ports = ["api"]
|
||||||
|
|
||||||
# Required for hardware access
|
|
||||||
privileged = true
|
privileged = true
|
||||||
|
|
||||||
# Pass the graphics hardware to the container
|
# --- THE FIX: STRICT MAPPING ---
|
||||||
|
# Only map the eGPU (renderD128) and the Compute interface (kfd)
|
||||||
|
devices = [
|
||||||
|
"/dev/kfd",
|
||||||
|
"/dev/dri/renderD128"
|
||||||
|
]
|
||||||
|
|
||||||
|
# Do NOT map the whole /dev/dri folder, or it might peek at the others
|
||||||
volumes = [
|
volumes = [
|
||||||
"/mnt/local-ssd/nomad/stacks/ai/ai-backend/ollama:/root/.ollama",
|
"/mnt/local-ssd/nomad/stacks/ai/ai-backend/ollama:/root/.ollama"
|
||||||
"/dev/kfd:/dev/kfd",
|
|
||||||
"/dev/dri:/dev/dri"
|
|
||||||
]
|
]
|
||||||
}
|
}
|
||||||
|
|
||||||
service {
|
service {
|
||||||
name = "ollama"
|
name = "ollama"
|
||||||
port = "api"
|
port = "api"
|
||||||
|
|||||||
Reference in New Issue
Block a user