fix(docker): default ROCm base 6.4.1 -> 7.2.2 (gfx1151 segfault on 7.x hosts)

Runtime verification on lucebox2 (gfx1151 Strix Halo, host ROCm 7.2.2):
the 6.4.1-userspace image finds the device but SIGSEGVs during model
load (and reports a bogus 1.28 TB VRAM total) — a 6.4.x-userspace /
7.x-host-driver mismatch. Rebuilding the same image with
ROCM_VERSION=7.2.2 works end-to-end: server up, /health + /props OK,
coherent chat completion at 12 tok/s decode on the iGPU.

Default both Dockerfile.rocm and docker-bake.hcl to 7.2.2 so the
published :rocm image runs on current ROCm 7.x host stacks; 6.4.x
remains available via the build arg for hosts still on a 6.x driver.

Co-Authored-By: WOZCODE <contact@withwoz.com>

Author: davide221

Dockerfile.rocm

@@ -3,13 +3,15 @@
 # ─── Stage 1: builder ───────────────────────────────────────────────────────
 # ROCm/HIP sibling of the CUDA Dockerfile. ROCM_VERSION / UBUNTU_VERSION /
 # DFLASH_HIP_ARCHES are build args so the same Dockerfile can be repinned.
-#   • lucebox-hub:rocm  — ROCm 6.4.x, gfx1151 (+ optional fat gfx list)
+#   • lucebox-hub:rocm  — ROCm 7.2.x, gfx1151 (+ optional fat gfx list)
 # See docker-bake.hcl for the canonical invocation.
 #
-# NOTE: gfx1151 (Strix Halo / Ryzen AI MAX) needs ROCm >= 6.4.1. The default
-# below is the earliest officially-published base with gfx1151 support; bump
-# to a 7.x tag (e.g. ROCM_VERSION=7.0.2) to match a newer host ROCm stack.
-ARG ROCM_VERSION=6.4.1
+# NOTE: gfx1151 (Strix Halo / Ryzen AI MAX) needs ROCm >= 6.4.1, but the
+# 6.4.x userspace segfaults at model load against a ROCm 7.x host driver
+# (verified on gfx1151 + host ROCm 7.2.2: SIGSEGV in backend creation, bogus
+# 1.28 TB VRAM report). Default to 7.2.2 to match current host stacks; pin
+# back to a 6.4.x tag only for hosts still on a 6.x driver.
+ARG ROCM_VERSION=7.2.2
 ARG UBUNTU_VERSION=22.04
 FROM rocm/dev-ubuntu-${UBUNTU_VERSION}:${ROCM_VERSION} AS builder
 

docker-bake.hcl

@@ -52,9 +52,11 @@ variable "DFLASH_CUDA_ARCHES" { default = "75;80;86;89;90;120" }
 # gfx90a MI200.)
 variable "DFLASH_HIP_ARCHES" { default = "gfx1151" }
 
-# ROCm base-image tag for the rocm variant. gfx1151 needs >= 6.4.1; bump to a
-# 7.x tag (e.g. "7.0.2") to match a newer host ROCm stack.
-variable "ROCM_VERSION" { default = "6.4.1" }
+# ROCm base-image tag for the rocm variant. gfx1151 needs >= 6.4.1, but the
+# 6.4.x userspace segfaults at model load against a ROCm 7.x host driver
+# (see Dockerfile.rocm); 7.2.2 matches current host stacks. Pin back to a
+# 6.4.x tag only for hosts still on a 6.x driver.
+variable "ROCM_VERSION" { default = "7.2.2" }
 
 # Image identity stamped into /opt/lucebox-hub/IMAGE_INFO at build time and
 # surfaced under /props.build at runtime (git_sha, image_tag, build_time).