From 4cd8d1df7d2382c8a2612172faa29f0dd8c92c4c Mon Sep 17 00:00:00 2001
From: worthant <b_dvorkin@niuitmo.ru>
Date: Fri, 24 Jul 2026 23:40:02 +0300
Subject: [PATCH] =?UTF-8?q?=F0=9F=94=A7=20ci(rocm):=20lean=20ROCm=20archiv?=
 =?UTF-8?q?e=20=E2=80=94=20stop=20bundling=20the=20multi-GB=20Tensile=20DB?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

The ROCm archive hit 9.6 GB and blew past GitHub's 2 GB asset limit, so
publish-release 422'd and b10018-1.2.0 failed to publish. Cause: I bundled
the whole rocBLAS/hipBLASLt Tensile kernel database (kernels for every gfx
arch) to make the archive self-contained. Drop it: ship binaries +
libggml-hip.so (~0.45 GB) only, require the system ROCm runtime like
upstream's ROCm builds. Archive drops ~9.6 GB -> ~0.6 GB. Adds a
README-ROCm.txt noting the runtime requirement.
---
 .github/workflows/dev-build.yml          | 20 ++++++++++----------
 .github/workflows/release-turboquant.yml | 20 ++++++++++----------
 2 files changed, 20 insertions(+), 20 deletions(-)

diff --git a/.github/workflows/dev-build.yml b/.github/workflows/dev-build.yml
index 0a4963c17a51..050bc8a9f2ec 100644
--- a/.github/workflows/dev-build.yml
+++ b/.github/workflows/dev-build.yml
@@ -275,10 +275,11 @@ jobs:
           ls -l build/bin/libggml-hip.so
 
       - name: Prepare archive
-        # Bundle the HIP runtime like the CUDA job bundles cudart/cublas:
-        # user boxes have the amdgpu driver but not necessarily ROCm
-        # userspace. rocBLAS/hipBLASLt also ship Tensile kernel data dirs
-        # that must travel with the .so or matmul fails at runtime.
+        # Lean archive: binaries + libggml-*.so only. Do NOT bundle the ROCm
+        # runtime — rocBLAS/hipBLASLt ship a Tensile kernel database for every
+        # gfx arch that balloons the archive past 9 GB (and GitHub's 2 GB asset
+        # limit). AMD users install the ROCm runtime system-wide, exactly like
+        # upstream's ROCm builds; libggml-hip.so dlopen-links it at runtime.
         run: |
           mkdir -p release/build/bin
           cp build/bin/llama-server release/build/bin/
@@ -287,13 +288,12 @@ jobs:
           cp build/bin/llama-perplexity release/build/bin/ 2>/dev/null || true
           cp build/bin/llama-quantize release/build/bin/ 2>/dev/null || true
           find build/bin -name "*.so*" -exec cp -P {} release/build/bin/ \; 2>/dev/null || true
-          for lib in libamdhip64 librocblas libhipblas libhipblaslt librocsolver librocsparse libamd_comgr libhsa-runtime64; do
-            cp -P /opt/rocm/lib/${lib}.so* release/build/bin/ 2>/dev/null || true
-          done
-          # Tensile / hipBLASLt kernel data (required at runtime by rocBLAS).
-          cp -r /opt/rocm/lib/rocblas release/build/bin/ 2>/dev/null || true
-          cp -r /opt/rocm/lib/hipblaslt release/build/bin/ 2>/dev/null || true
           cp LICENSE release/build/bin/ 2>/dev/null || true
+          cat > release/build/bin/README-ROCm.txt << 'EOF'
+          This build needs the AMD ROCm runtime installed on the system
+          (https://rocm.docs.amd.com). Targets AMD RDNA2-RDNA4:
+          gfx1030/1100/1101/1102/1151/1200/1201. GCN GPUs: use the Vulkan build.
+          EOF
           cd release
           zip -r ../llama-turboquant-linux-x64-rocm.zip .
           tar -czf ../llama-turboquant-linux-x64-rocm.tar.gz .
diff --git a/.github/workflows/release-turboquant.yml b/.github/workflows/release-turboquant.yml
index 272b27125720..6f75b836cb64 100644
--- a/.github/workflows/release-turboquant.yml
+++ b/.github/workflows/release-turboquant.yml
@@ -280,10 +280,11 @@ jobs:
           ls -l build/bin/libggml-hip.so
 
       - name: Prepare archive
-        # Bundle the HIP runtime like the CUDA job bundles cudart/cublas:
-        # user boxes have the amdgpu driver but not necessarily ROCm
-        # userspace. rocBLAS/hipBLASLt also ship Tensile kernel data dirs
-        # that must travel with the .so or matmul fails at runtime.
+        # Lean archive: binaries + libggml-*.so only. Do NOT bundle the ROCm
+        # runtime — rocBLAS/hipBLASLt ship a Tensile kernel database for every
+        # gfx arch that balloons the archive past 9 GB (and GitHub's 2 GB asset
+        # limit). AMD users install the ROCm runtime system-wide, exactly like
+        # upstream's ROCm builds; libggml-hip.so dlopen-links it at runtime.
         run: |
           mkdir -p release/build/bin
           cp build/bin/llama-server release/build/bin/
@@ -292,13 +293,12 @@ jobs:
           cp build/bin/llama-perplexity release/build/bin/ 2>/dev/null || true
           cp build/bin/llama-quantize release/build/bin/ 2>/dev/null || true
           find build/bin -name "*.so*" -exec cp -P {} release/build/bin/ \; 2>/dev/null || true
-          for lib in libamdhip64 librocblas libhipblas libhipblaslt librocsolver librocsparse libamd_comgr libhsa-runtime64; do
-            cp -P /opt/rocm/lib/${lib}.so* release/build/bin/ 2>/dev/null || true
-          done
-          # Tensile / hipBLASLt kernel data (required at runtime by rocBLAS).
-          cp -r /opt/rocm/lib/rocblas release/build/bin/ 2>/dev/null || true
-          cp -r /opt/rocm/lib/hipblaslt release/build/bin/ 2>/dev/null || true
           cp LICENSE release/build/bin/ 2>/dev/null || true
+          cat > release/build/bin/README-ROCm.txt << 'EOF'
+          This build needs the AMD ROCm runtime installed on the system
+          (https://rocm.docs.amd.com). Targets AMD RDNA2-RDNA4:
+          gfx1030/1100/1101/1102/1151/1200/1201. GCN GPUs: use the Vulkan build.
+          EOF
           cd release
           zip -r ../llama-turboquant-linux-x64-rocm.zip .
           tar -czf ../llama-turboquant-linux-x64-rocm.tar.gz .