feat: sglang backend support

10 files changed, 80 insertions, 0 deletions

diff --git a/app/backends/sglomni/configs/dots_tts.yaml b/app/backends/sglomni/configs/dots_tts.yaml
new file mode 100644
index 0000000..d2c42f4
--- /dev/null
+++ b/app/backends/sglomni/configs/dots_tts.yaml
@@ -0,0 +1,36 @@
+config_cls: DotsTTSPipelineConfig
+model_path: dots-studio/dots.tts-mf
+
+# The solver settings feed two consumers: the latent AR engine solves with
+# them, and preprocessing sizes its generation schedule from them. The
+# selector writes one value into both stages; an explicit per-stage entry
+# under stages: would override it.
+shared:
+  - select:
+      stages: [preprocessing, latent_engine]
+    factory:
+      num_steps: 4
+      max_generate_length: 500
+
+stages:
+  reference_encode:
+    factory:
+      max_concurrency: 8
+      max_batch_size: 1
+      max_batch_wait_ms: 4
+  latent_engine:
+    factory:
+      optimize: true
+    engine:
+      mem_fraction_static: 0.20
+      max_running_requests: 16
+      # note (luojiaxuan): backbone decode runs through the SGLang CUDA graph
+      # with the model-owned feedback buffer; delete these two lines (or set
+      # disable_cuda_graph: true) to fall back to eager backbone decode.
+      disable_cuda_graph: false
+      cuda_graph_max_bs: 16
+  vocoder:
+    factory:
+      optimize: true
+      max_batch_size: 4
+      max_batch_wait_ms: 2
diff --git a/app/backends/sglomni/configs/moss_tts.yaml b/app/backends/sglomni/configs/moss_tts.yaml
new file mode 100644
index 0000000..ef8d8fa
--- /dev/null
+++ b/app/backends/sglomni/configs/moss_tts.yaml
@@ -0,0 +1,2 @@
+config_cls: MossTTSPipelineConfig
+model_path: OpenMOSS-Team/MOSS-TTS-v1.5
diff --git a/app/backends/sglomni/configs/moss_tts_local.yaml b/app/backends/sglomni/configs/moss_tts_local.yaml
new file mode 100644
index 0000000..0b37f4e
--- /dev/null
+++ b/app/backends/sglomni/configs/moss_tts_local.yaml
@@ -0,0 +1,2 @@
+config_cls: MossTTSLocalPipelineConfig
+model_path: OpenMOSS-Team/MOSS-TTS-Local-Transformer-v1.5
diff --git a/app/backends/sglomni/configs/qwen3_tts_0_6b.yaml b/app/backends/sglomni/configs/qwen3_tts_0_6b.yaml
new file mode 100644
index 0000000..a712ef9
--- /dev/null
+++ b/app/backends/sglomni/configs/qwen3_tts_0_6b.yaml
@@ -0,0 +1,2 @@
+config_cls: Qwen3TTSPipelineConfig
+model_path: Qwen/Qwen3-TTS-12Hz-0.6B-Base
diff --git a/app/backends/sglomni/configs/qwen3_tts_0_6b_customvoice.yaml b/app/backends/sglomni/configs/qwen3_tts_0_6b_customvoice.yaml
new file mode 100644
index 0000000..6b284da
--- /dev/null
+++ b/app/backends/sglomni/configs/qwen3_tts_0_6b_customvoice.yaml
@@ -0,0 +1,2 @@
+config_cls: Qwen3TTSPipelineConfig
+model_path: Qwen/Qwen3-TTS-12Hz-0.6B-CustomVoice
diff --git a/app/backends/sglomni/configs/qwen3_tts_1_7b.yaml b/app/backends/sglomni/configs/qwen3_tts_1_7b.yaml
new file mode 100644
index 0000000..4f7706d
--- /dev/null
+++ b/app/backends/sglomni/configs/qwen3_tts_1_7b.yaml
@@ -0,0 +1,2 @@
+config_cls: Qwen3TTSPipelineConfig
+model_path: Qwen/Qwen3-TTS-12Hz-1.7B-Base
diff --git a/app/backends/sglomni/configs/qwen3_tts_1_7b_voicedesign.yaml b/app/backends/sglomni/configs/qwen3_tts_1_7b_voicedesign.yaml
new file mode 100644
index 0000000..20ae0b8
--- /dev/null
+++ b/app/backends/sglomni/configs/qwen3_tts_1_7b_voicedesign.yaml
@@ -0,0 +1,2 @@
+config_cls: Qwen3TTSPipelineConfig
+model_path: Qwen/Qwen3-TTS-12Hz-1.7B-VoiceDesign
diff --git a/app/backends/sglomni/configs/s2pro_tts.yaml b/app/backends/sglomni/configs/s2pro_tts.yaml
new file mode 100644
index 0000000..0bc3eba
--- /dev/null
+++ b/app/backends/sglomni/configs/s2pro_tts.yaml
@@ -0,0 +1,2 @@
+config_cls: S2ProPipelineConfig
+model_path: fishaudio/s2-pro
diff --git a/app/backends/sglomni/configs/voxtral_tts.yaml b/app/backends/sglomni/configs/voxtral_tts.yaml
new file mode 100644
index 0000000..450cbce
--- /dev/null
+++ b/app/backends/sglomni/configs/voxtral_tts.yaml
@@ -0,0 +1,2 @@
+config_cls: VoxtralTTSPipelineConfig
+model_path: mistralai/Voxtral-4B-TTS-2603
diff --git a/app/backends/sglomni/configs/zonos2_bf16.yaml b/app/backends/sglomni/configs/zonos2_bf16.yaml
new file mode 100644
index 0000000..79d5d26
--- /dev/null
+++ b/app/backends/sglomni/configs/zonos2_bf16.yaml
@@ -0,0 +1,28 @@
+# ZONOS2 without the FP8-quantized MoE pipeline.
+#
+# sglang-omni's default ZONOS2 config (Zonos2PipelineConfig) hardcodes
+# `fp8: True` in the tts_engine stage factory: the MoE experts are
+# dynamically quantized bf16 -> fp8 at load time, and sglang's fused-MoE
+# Triton kernel for fp8e4nv only compiles on compute capability 8.9+
+# (RTX 4090/5090, Hopper). On older GPUs the server dies mid-boot with
+# `type fp8e4nv not supported in this architecture`.
+#
+# This copy turns FP8 off (`factory.fp8: false` overrides the hardcoded
+# kwarg — free-form factory keys pass through to the stage factory), so
+# the model runs in bf16 and works on e.g. Ampere (RTX 30xx, A100) at
+# roughly twice the MoE VRAM. The managed spec selects it automatically
+# on GPUs the FP8 path cannot run (backends.sglomni.status.build_spec).
+#
+# The bf16 weights (~11.5 GB) also need a bigger static-pool budget than
+# the builder's default 0.5: on a 24 GB card 0.5 leaves no room for the
+# KV cache inside 12 GB, and the server aborts with "Loaded weights
+# leave no GPU memory for the KV cache" (the profiler asks for >=0.64).
+# 0.70 gives ~16.8 GB static -> ~1.4 GB KV, and leaves ~7 GB of the card
+# for the colocated speaker-encode/vocoder stages.
+config_cls: Zonos2PipelineConfig
+model_path: Zyphra/zonos2
+stages:
+  tts_engine:
+    factory:
+      fp8: false
+      mem_fraction_static: 0.70