feat: sglang backend support

1 files changed, 36 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