fix(docker): keep ROCm base default at 6.4.1 (7.2.x intermittent on Strix Halo)

Revert the 7.2.2 default: the 7.2.x stack has shown intermittent
problems on Strix Halo, so the published :rocm image stays on 6.4.1.
The 6.4.x-userspace / 7.x-host-driver segfault from the previous commit
remains documented in Dockerfile.rocm + docker-bake.hcl: on a ROCm 7.x
host, rebuild with ROCM_VERSION=7.2.2 to match the host driver.

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

Author: davide221

Dockerfile.rocm

@@ -3,15 +3,17 @@
 # ─── 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 7.2.x, gfx1151 (+ optional fat gfx list)
+#   • lucebox-hub:rocm  — ROCm 6.4.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, 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
+# NOTE: gfx1151 (Strix Halo / Ryzen AI MAX) needs ROCm >= 6.4.1. The default
+# stays on 6.4.1 because the 7.2.x stack has shown intermittent problems on
+# Strix Halo. The flip side: against a ROCm 7.x HOST driver the 6.4.x
+# userspace can segfault at model load (seen on gfx1151 + host ROCm 7.2.2:
+# SIGSEGV in backend creation, bogus 1.28 TB VRAM report) — in that case
+# rebuild with ROCM_VERSION=7.2.2 to match the host. Rule of thumb: keep the
+# base's major version aligned with the host driver.
+ARG ROCM_VERSION=6.4.1
 ARG UBUNTU_VERSION=22.04
 FROM rocm/dev-ubuntu-${UBUNTU_VERSION}:${ROCM_VERSION} AS builder
 

docker-bake.hcl

@@ -52,11 +52,12 @@ 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, 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" }
+# ROCm base-image tag for the rocm variant. gfx1151 needs >= 6.4.1. Default
+# stays 6.4.1 (7.2.x has shown intermittent problems on Strix Halo), but on a
+# ROCm 7.x HOST driver the 6.4.x userspace can segfault at model load — set
+# ROCM_VERSION=7.2.2 there. Keep the base aligned with the host driver (see
+# Dockerfile.rocm).
+variable "ROCM_VERSION" { default = "6.4.1" }
 
 # 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).