From 7d0f799e1848c9de9ffd5fdddb34fc83135ed8cf Mon Sep 17 00:00:00 2001
From: mudler <2420543+mudler@users.noreply.github.com>
Date: Sun, 25 Oct 2026 21:31:04 +0000
Subject: [PATCH 1/4] :arrow_up: Update NVIDIA/NeMo-Speech.cpp

Signed-off-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>
---
 backend/go/nemo-speech-cpp/Makefile | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/backend/go/nemo-speech-cpp/Makefile b/backend/go/nemo-speech-cpp/Makefile
index 9d76b91f7fbe..7f0b81619c2e 100644
--- a/backend/go/nemo-speech-cpp/Makefile
+++ b/backend/go/nemo-speech-cpp/Makefile
@@ -12,7 +12,7 @@
 # runs 'make -C backend/go/$(BACKEND) build' and then copies package/), so it
 # has to produce the binary and the package, not just the shared libraries.
 
-NEMO_SPEECH_VERSION?=4c101bc7113f49101a3e11d2c994c519f41939f6
+NEMO_SPEECH_VERSION?=e680fc30ac7404ba3acd58532f713becf8625894
 NEMO_SPEECH_REPO?=https://github.com/NVIDIA/NeMo-Speech.cpp
 
 GOCMD?=go

From 5c70d20e5e202efcccbc21582ff4a5e13db581a6 Mon Sep 17 00:00:00 2001
From: Ettore Di Giacinto <mudler@localai.io>
Date: Mon, 26 Oct 2026 09:38:11 +0000
Subject: [PATCH 2/4] fix(nemo-speech-cpp): use upstream llama.cpp ggml
 patching

Initialize ggml through the llama.cpp submodule and let upstream CMake apply the patch series with fail-closed error handling.

Assisted-by: nib:gpt-6.1-sol
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
---
 backend/go/nemo-speech-cpp/Makefile | 33 ++++++++---------------------
 1 file changed, 9 insertions(+), 24 deletions(-)

diff --git a/backend/go/nemo-speech-cpp/Makefile b/backend/go/nemo-speech-cpp/Makefile
index 7f0b81619c2e..2de10ebdbbe2 100644
--- a/backend/go/nemo-speech-cpp/Makefile
+++ b/backend/go/nemo-speech-cpp/Makefile
@@ -152,11 +152,8 @@ endif
 # behind NEMO_SPEECH_GGML_PATCHED itself, which guards a Q8_PLANAR flag write
 # that a non-CUDA buffer already throws before reaching.
 #
-# So on macOS the series buys nothing, and it cannot be applied there anyway:
-# upstream's scripts/apply-ggml-patches.sh uses mapfile, a bash 4 builtin, and
-# macOS ships bash 3.2 as the only bash on the runner's PATH. Skip the patch
-# step and tell cmake the linked ggml is stock, which is exactly upstream's own
-# Metal configuration. Linux keeps applying the series unchanged.
+# On macOS keep ggml stock, matching upstream's Metal configuration. Upstream
+# CMake now manages the patch series; Linux keeps enabling it by default.
 ifeq ($(shell uname -s),Darwin)
 	GGML_PATCHED?=OFF
 else
@@ -175,28 +172,16 @@ sources/NeMo-Speech.cpp:
 	# NMT links llama.cpp; ja needs open_jtalk; zh needs cppjieba. flashlight and
 	# kenlm are deliberately not initialized, they are out of scope.
 	cd sources/NeMo-Speech.cpp && git submodule update --init --recursive \
-		ggml llama.cpp third_party/open_jtalk third_party/cppjieba third_party/cpp-httplib
+		llama.cpp third_party/open_jtalk third_party/cppjieba third_party/cpp-httplib
 
-# NEMO_SPEECH_GGML_PATCHED defaults ON and silently assumes the ggml-patches
-# series is applied. An unpatched checkout builds fine and produces wrong CUDA
-# encoder output, so a failure here must stop the build rather than warn.
+# ggml now comes from llama.cpp. Upstream's cmake/llama_cpp.cmake applies
+# patches/series to a build-directory copy when NEMO_SPEECH_GGML_PATCHED=ON
+# and fails configuration if a patch cannot apply.
 #
-# Upstream's own script is the right tool: it applies the series in filename
-# order, exits non-zero when a patch does not apply, and decides "already
-# applied" by comparing the full-series tree hash rather than a timestamp. That
-# makes it safe to run unconditionally, so there is no sentinel file to go stale
-# or to wedge the build when deleted.
-#
-# Both branches keep the order-only clone prerequisite: it is the only thing
-# that pulls sources/ in on a WITH_NORM=OFF tree, where the library rule has no
-# other prerequisite left.
-ifeq ($(GGML_PATCHED),ON)
-patch-ggml: | sources/NeMo-Speech.cpp
-	cd sources/NeMo-Speech.cpp && bash scripts/apply-ggml-patches.sh
-else
+# Keep the order-only clone prerequisite so patch-ggml still initializes sources
+# before the runtime's CMake configuration.
 patch-ggml: | sources/NeMo-Speech.cpp
-	@echo "[ggml-patch] skipped: NEMO_SPEECH_GGML_PATCHED=$(GGML_PATCHED), the series is CUDA-only"
-endif
+	@echo "[ggml-patch] managed by upstream CMake (NEMO_SPEECH_GGML_PATCHED=$(GGML_PATCHED))"
 
 # The Sparrowhawk/OpenFST text-normalization stack, as a target in its own right
 # keyed on a file the build script actually produces.

From e1232e279a47727eeaaca8e4ca5a86942d54ee83 Mon Sep 17 00:00:00 2001
From: Ettore Di Giacinto <mudler@localai.io>
Date: Mon, 26 Oct 2026 14:26:19 +0000
Subject: [PATCH 3/4] fix(nemo-speech-cpp): recognize upstream OpenFST
 ownership fix

Accept both WrapUnique assignments unchanged while preserving the legacy header patch and rejecting unknown or incomplete forms.

Assisted-by: nib:gpt-6.1-sol
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
---
 backend/go/nemo-speech-cpp/Makefile | 32 +++++++++++++++++------------
 1 file changed, 19 insertions(+), 13 deletions(-)

diff --git a/backend/go/nemo-speech-cpp/Makefile b/backend/go/nemo-speech-cpp/Makefile
index 2de10ebdbbe2..d39610cf73f3 100644
--- a/backend/go/nemo-speech-cpp/Makefile
+++ b/backend/go/nemo-speech-cpp/Makefile
@@ -242,23 +242,29 @@ $(ITN_MARKER): | sources/NeMo-Speech.cpp
 # compiler could ever have emitted it. Only these two lines are affected: gcc 14
 # reports exactly two errors over the whole OpenFST include closure, both here.
 #
-# Guarded on both sides so a pinned-version bump cannot silently no-op it: the
-# first check fails if neither the broken nor the fixed spelling is present, the
-# last fails if the broken one survives.
+# The newer OpenFST pin fixes both assignments with fst::WrapUnique. Accept
+# that form unchanged; otherwise require both legacy assignments to be known
+# before patching, and reject unknown or incomplete forms rather than no-op.
 patch-itn-headers:
 	@test -f $(ITN_FST_HEADER) || { \
 		echo "ERROR: $(ITN_FST_HEADER) missing; the ITN prefix is not installed." >&2; \
 		exit 1; }
-	@grep -q 'isymbols_ = impl.isymbols_' $(ITN_FST_HEADER) || \
-	 grep -q 'isymbols_.reset(impl.isymbols_' $(ITN_FST_HEADER) || { \
-		echo "ERROR: FstImpl::operator= in $(ITN_FST_HEADER) matches neither the" >&2; \
-		echo "       known-broken nor the patched form. OpenFST changed upstream;" >&2; \
-		echo "       re-check whether this patch is still needed before removing it." >&2; \
-		exit 1; }
-	sed -i -E 's|^([[:space:]]*)([io]symbols_) = (impl\.[io]symbols_ \? impl\.[io]symbols_->Copy\(\) : nullptr);$$|\1\2.reset(\3);|' $(ITN_FST_HEADER)
-	@if grep -q 'symbols_ = impl.[io]symbols_' $(ITN_FST_HEADER); then \
-		echo "ERROR: the FstImpl::operator= patch did not apply to $(ITN_FST_HEADER)." >&2; \
-		exit 1; \
+	@if grep -Fq 'isymbols_ = fst::WrapUnique(impl.isymbols_ ? impl.isymbols_->Copy() : nullptr);' $(ITN_FST_HEADER) && \
+	    grep -Fq 'osymbols_ = fst::WrapUnique(impl.osymbols_ ? impl.osymbols_->Copy() : nullptr);' $(ITN_FST_HEADER); then \
+		echo "OpenFST copy assignment already fixed upstream."; \
+	else \
+		for symbol in isymbols_ osymbols_; do \
+			grep -Fq "$$symbol = impl.$$symbol ? impl.$$symbol->Copy() : nullptr;" $(ITN_FST_HEADER) || \
+			grep -Fq "$$symbol.reset(impl.$$symbol ? impl.$$symbol->Copy() : nullptr);" $(ITN_FST_HEADER) || { \
+				echo "ERROR: FstImpl::operator= in $(ITN_FST_HEADER) has an unknown $$symbol assignment." >&2; \
+				echo "       OpenFST changed upstream; re-check whether this patch is still needed." >&2; \
+				exit 1; }; \
+		done; \
+		sed -i -E 's|^([[:space:]]*)([io]symbols_) = (impl\.[io]symbols_ \? impl\.[io]symbols_->Copy\(\) : nullptr);$$|\1\2.reset(\3);|' $(ITN_FST_HEADER) || exit 1; \
+		if grep -q 'symbols_ = impl.[io]symbols_' $(ITN_FST_HEADER); then \
+			echo "ERROR: the FstImpl::operator= patch did not apply to $(ITN_FST_HEADER)." >&2; \
+			exit 1; \
+		fi; \
 	fi
 
 itn: $(ITN_MARKER)

From ccd4a357eef1a4d701abe55704e01c3532c587b0 Mon Sep 17 00:00:00 2001
From: Ettore Di Giacinto <mudler@localai.io>
Date: Mon, 26 Oct 2026 16:12:18 +0000
Subject: [PATCH 4/4] fix(nemo-speech-cpp): bundle shared text normalization
 runtime

Stage the normalization DSO when ITN is present and include it in the package alongside its dependency closure. Fail staging if the required DSO is missing.

Assisted-by: nib:gpt-6.1-sol
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
---
 backend/go/nemo-speech-cpp/Makefile   | 8 ++++----
 backend/go/nemo-speech-cpp/package.sh | 9 ++++-----
 2 files changed, 8 insertions(+), 9 deletions(-)

diff --git a/backend/go/nemo-speech-cpp/Makefile b/backend/go/nemo-speech-cpp/Makefile
index d39610cf73f3..206a5813d1a9 100644
--- a/backend/go/nemo-speech-cpp/Makefile
+++ b/backend/go/nemo-speech-cpp/Makefile
@@ -320,10 +320,9 @@ stage-libs: sources/NeMo-Speech.cpp/build/bin/libnemo_speech_asr_c.so
 	# no counterpart, its implementation and ABI live in the same object.
 	cp -af sources/NeMo-Speech.cpp/build/bin/libnemo_speech_asr.* .
 	cp -af sources/NeMo-Speech.cpp/build/bin/libnemo_speech_nmt.* .
-	# nemo_speech_text_normalization is STATIC but links sparrowhawk, fstfar and
-	# fst PUBLIC, so those become DT_NEEDED on libnemo_speech_asr.so. They live in
-	# a project-local prefix that nothing else on the system provides, so without
-	# staging them here the packaged backend cannot dlopen at all.
+	# ASR and TTS now share nemo_speech_text_normalization as a separate DSO.
+	# Stage it alongside its Sparrowhawk/OpenFST dependencies from the local
+	# prefix so the packaged backend can resolve the whole normalization stack.
 	#
 	# Keyed on the prefix existing rather than on WITH_NORM, so this stages what
 	# the tree actually built. A WITH_NORM=ON build cannot reach here without the
@@ -331,6 +330,7 @@ stage-libs: sources/NeMo-Speech.cpp/build/bin/libnemo_speech_asr_c.so
 	# library that needs Sparrowhawk somehow arrives unstaged, package.sh's
 	# closure guard fails the build rather than shipping it.
 	@if [ -d "$(ITN_LIB_DIR)" ]; then \
+		cp -af sources/NeMo-Speech.cpp/build/bin/libnemo_speech_text_normalization.* . || exit 1; \
 		echo "cp -af $(ITN_LIB_DIR)/*.so* ."; \
 		cp -af $(ITN_LIB_DIR)/*.so* .; \
 	fi
diff --git a/backend/go/nemo-speech-cpp/package.sh b/backend/go/nemo-speech-cpp/package.sh
index c0cc3f965b8b..7051e09487af 100755
--- a/backend/go/nemo-speech-cpp/package.sh
+++ b/backend/go/nemo-speech-cpp/package.sh
@@ -40,14 +40,13 @@ for lib in libnemo_speech_asr_c libnemo_speech_asr libnemo_speech_tts libnemo_sp
 	fi
 done
 
-# Text normalization (WITH_NORM=ON, Linux only) links Sparrowhawk and OpenFST
-# into libnemo_speech_asr.so. Those live in a project-local prefix that the
-# Makefile stages here, so anything staged that is not a nemo_speech object is
-# part of that stack. Absent on a WITH_NORM=OFF build, which is why this is a
-# glob that tolerates no matches rather than a required list.
+# Text normalization (WITH_NORM=ON, Linux only) is a shared NeMo library with
+# Sparrowhawk/OpenFST dependencies. Bundle it and the prefix libraries staged
+# by the Makefile. Absent on a WITH_NORM=OFF build, so tolerate no matches.
 shopt -s nullglob
 for so in "$CURDIR"/*.so "$CURDIR"/*.so.* "$CURDIR"/*.dylib; do
 	case "$(basename "$so")" in
+		libnemo_speech_text_normalization.*) ;;
 		libnemo_speech_*) continue ;;
 	esac
 	cp -avf "$so" "$CURDIR/package/lib/"