From 41d4251cd5cdfa19b63275848fc5da9f29a1bbbf Mon Sep 17 00:00:00 2001
From: localai-org-maint-bot
 <306269227+localai-org-maint-bot@users.noreply.github.com>
Date: Fri, 31 Jul 2026 12:04:21 +0000
Subject: [PATCH 1/4] fix(qwen-tts): install flash-attn on CUDA 13

Add the missing cublas13 post-install requirements hook so the CUDA 13 backend uses FlashAttention instead of falling back to SDPA. Keep CUDA packaging parity covered by the backend test script.

Assisted-by: Codex:gpt-5
---
 backend/python/qwen-tts/requirements-cublas13-after.txt | 1 +
 backend/python/qwen-tts/test.sh                         | 5 +++++
 2 files changed, 6 insertions(+)
 create mode 100644 backend/python/qwen-tts/requirements-cublas13-after.txt

diff --git a/backend/python/qwen-tts/requirements-cublas13-after.txt b/backend/python/qwen-tts/requirements-cublas13-after.txt
new file mode 100644
index 000000000000..d0f50993616a
--- /dev/null
+++ b/backend/python/qwen-tts/requirements-cublas13-after.txt
@@ -0,0 +1 @@
+flash-attn
diff --git a/backend/python/qwen-tts/test.sh b/backend/python/qwen-tts/test.sh
index eb59f2aaf3f3..e4833fe5e627 100755
--- a/backend/python/qwen-tts/test.sh
+++ b/backend/python/qwen-tts/test.sh
@@ -2,6 +2,11 @@
 set -e
 
 backend_dir=$(dirname $0)
+
+for cuda_version in 12 13; do
+    grep -qx "flash-attn" "$backend_dir/requirements-cublas${cuda_version}-after.txt"
+done
+
 if [ -d $backend_dir/common ]; then
     source $backend_dir/common/libbackend.sh
 else

From abb6e2b9d02f14d35da6185da92188b33d5e4c0e Mon Sep 17 00:00:00 2001
From: localai-org-maint-bot
 <306269227+localai-org-maint-bot@users.noreply.github.com>
Date: Sat, 1 Aug 2026 00:06:44 +0000
Subject: [PATCH 2/4] fix(qwen-tts): serialize CUDA 13 flash-attn build

Limit the CUDA 13 FlashAttention source build to one ninja worker by default so it stays within CI runner memory. Preserve explicit MAX_JOBS overrides and leave other build profiles unchanged.

Assisted-by: Codex:gpt-5 [web]
---
 backend/python/qwen-tts/install.sh | 6 ++++++
 1 file changed, 6 insertions(+)

diff --git a/backend/python/qwen-tts/install.sh b/backend/python/qwen-tts/install.sh
index b7d487873f78..668969fff9fc 100755
--- a/backend/python/qwen-tts/install.sh
+++ b/backend/python/qwen-tts/install.sh
@@ -10,4 +10,10 @@ else
     source $backend_dir/../common/libbackend.sh
 fi
 
+# CUDA 13 has no prebuilt FlashAttention wheel, so the fallback source build
+# exceeds the CI runner's memory when ninja compiles multiple units at once.
+if [ "x${BUILD_PROFILE}" = "xcublas13" ]; then
+    export MAX_JOBS="${MAX_JOBS:-1}"
+fi
+
 installRequirements

From 7916c48f04fc0c6c6d61a2651201b3c06d332ae0 Mon Sep 17 00:00:00 2001
From: localai-org-maint-bot
 <306269227+localai-org-maint-bot@users.noreply.github.com>
Date: Sat, 1 Aug 2026 03:03:20 +0000
Subject: [PATCH 3/4] test(qwen-tts): guard CUDA 13 build concurrency

Keep the CUDA 13 FlashAttention packaging check tied to its serialized source-build guard so the runner OOM fix cannot be removed unnoticed.\n\nAssisted-by: Codex:gpt-5 [Codex]
---
 backend/python/qwen-tts/test.sh | 3 +++
 1 file changed, 3 insertions(+)

diff --git a/backend/python/qwen-tts/test.sh b/backend/python/qwen-tts/test.sh
index e4833fe5e627..df31e076533a 100755
--- a/backend/python/qwen-tts/test.sh
+++ b/backend/python/qwen-tts/test.sh
@@ -7,6 +7,9 @@ for cuda_version in 12 13; do
     grep -qx "flash-attn" "$backend_dir/requirements-cublas${cuda_version}-after.txt"
 done
 
+grep -q 'BUILD_PROFILE.*cublas13' "$backend_dir/install.sh"
+grep -q 'MAX_JOBS.*1' "$backend_dir/install.sh"
+
 if [ -d $backend_dir/common ]; then
     source $backend_dir/common/libbackend.sh
 else

From 127a5780f20dde5ba0c83dbb972fbee87fe6993d Mon Sep 17 00:00:00 2001
From: localai-org-maint-bot
 <306269227+localai-org-maint-bot@users.noreply.github.com>
Date: Sat, 1 Aug 2026 20:03:17 +0000
Subject: [PATCH 4/4] fix(qwen-tts): limit CUDA compiler threads

The CUDA 13 FlashAttention build still exhausts hosted-runner memory with a single ninja worker because nvcc can compile multiple threads internally. Limit nvcc to one thread for that profile and guard the setting in the backend test script.

Assisted-by: Codex:gpt-5
---
 backend/python/qwen-tts/install.sh | 1 +
 backend/python/qwen-tts/test.sh    | 1 +
 2 files changed, 2 insertions(+)

diff --git a/backend/python/qwen-tts/install.sh b/backend/python/qwen-tts/install.sh
index 668969fff9fc..9c8bcf83f39c 100755
--- a/backend/python/qwen-tts/install.sh
+++ b/backend/python/qwen-tts/install.sh
@@ -14,6 +14,7 @@ fi
 # exceeds the CI runner's memory when ninja compiles multiple units at once.
 if [ "x${BUILD_PROFILE}" = "xcublas13" ]; then
     export MAX_JOBS="${MAX_JOBS:-1}"
+    export NVCC_THREADS="${NVCC_THREADS:-1}"
 fi
 
 installRequirements
diff --git a/backend/python/qwen-tts/test.sh b/backend/python/qwen-tts/test.sh
index df31e076533a..97ea2faf916d 100755
--- a/backend/python/qwen-tts/test.sh
+++ b/backend/python/qwen-tts/test.sh
@@ -9,6 +9,7 @@ done
 
 grep -q 'BUILD_PROFILE.*cublas13' "$backend_dir/install.sh"
 grep -q 'MAX_JOBS.*1' "$backend_dir/install.sh"
+grep -q 'NVCC_THREADS.*1' "$backend_dir/install.sh"
 
 if [ -d $backend_dir/common ]; then
     source $backend_dir/common/libbackend.sh