MiniMax Music3 preview: fix BF16 conv corruption, repair converter, wire web UI

include/engine/community_models/minimax_music3/assets.h

@@ -24,4 +24,14 @@ std::shared_ptr<const MiniMaxMusic3Assets> load_minimax_music3_assets(
     const std::filesystem::path & model_path);
 void validate_minimax_music3_anchors(const MiniMaxMusic3Assets & assets);
 
+// Conv kernels must not be stored BF16: the CUDA im2col path asserts on BF16 kernel
+// types in its naive lowering and silently corrupts through the fast path, which turns
+// the whole flow output into clipped noise. Conv weights are a negligible share of the
+// package, so BF16 requests (explicit, or Native over a BF16 source tensor) fall back
+// to F32 for convolutions.
+assets::TensorStorageType conv_safe_storage_type(
+    const assets::TensorSource & source,
+    const std::string & tensor_prefix,
+    assets::TensorStorageType requested);
+
 }  // namespace engine::models::minimax_music3

scripts/minimax_music3/convert_gguf.py

@@ -34,6 +34,24 @@
     ("vocoder", "vocoder/diffusion_pytorch_model.safetensors", "vocoder.gguf"),
 ]
 
+# Conv kernels must not be stored BF16: the CUDA conv lowering corrupts on BF16 kernel
+# types (and the runtime guards this at load). Keep them F32 on disk for every target
+# type.
+KEEP_TYPES = {
+    "transformer": ["preprocess_conv*=f32", "postprocess_conv*=f32"],
+    "condition_encoder": ["proj.weight=f32"],
+    "vocoder": ["dec_in_proj*=f32"],
+}
+
+# Config sidecars the runtime requires next to the component GGUFs.
+CONFIG_SIDECARS = {
+    "language_model": "language_model/config.json",
+    "rvq_depth_decoder": "rvq_depth_decoder/config.json",
+    "condition_encoder": "condition_encoder/config.json",
+    "transformer": "transformer/config.json",
+    "vocoder": "vocoder/config.json",
+}
+
 
 def parse_args() -> argparse.Namespace:
     parser = argparse.ArgumentParser(description=__doc__)
@@ -117,6 +135,7 @@ def write_variant_spec(args: argparse.Namespace) -> tempfile.TemporaryDirectory[
 
     output_by_name = {name: output_name(args, component) for name, _input, _base in COMPONENTS for component in [next(c for c in COMPONENTS if c[0] == name)]}
     gguf_source = next(source for source in spec["sources"] if source["format"] == "gguf")
+    gguf_source.setdefault("tensors", {})
     gguf_source["tensors"]["language_model_weights"] = f"model:{output_by_name['language_model']}"
     gguf_source["tensors"]["depth_decoder_weights"] = f"model:{output_by_name['rvq_depth_decoder']}"
     gguf_source["tensors"]["condition_encoder_weights"] = f"model:{output_by_name['condition_encoder']}"
@@ -151,11 +170,22 @@ def run_conversion(args: argparse.Namespace, component: tuple[str, str, str]) ->
         "--model-spec",
         str(args.model_spec),
         "--no-sidecars",
+        # The runtime opens component GGUFs by filename rather than through spec tensor
+        # ids, and the spec's gguf source does not yet match audiocpp_gguf's
+        # source-matching rules, so conversion proceeds without a matching package spec.
+        "--allow-missing-model-spec",
     ]
+    for keep in KEEP_TYPES.get(name, []):
+        cmd.extend(["--keep-type", keep])
     if args.overwrite:
         cmd.append("--overwrite")
     print("[convert]", name, "->", output_path, flush=True)
     subprocess.run(cmd, cwd=REPO_ROOT, check=True)
+    sidecar = CONFIG_SIDECARS.get(name)
+    if sidecar is not None:
+        config_dir = Path(args.output) / "config"
+        config_dir.mkdir(parents=True, exist_ok=True)
+        shutil.copyfile(Path(args.source) / sidecar, config_dir / f"{name}.json")
 
 
 def main() -> None:

src/community_models/minimax_music3/assets.cpp

@@ -147,4 +147,19 @@ std::shared_ptr<const MiniMaxMusic3Assets> load_minimax_music3_assets(
     return assets;
 }
 
+assets::TensorStorageType conv_safe_storage_type(
+    const assets::TensorSource & source,
+    const std::string & tensor_prefix,
+    assets::TensorStorageType requested) {
+    if (requested == assets::TensorStorageType::BF16) {
+        return assets::TensorStorageType::F32;
+    }
+    if (requested == assets::TensorStorageType::Native &&
+        assets::ggml_type_for_tensor_dtype(source.require_metadata(tensor_prefix + ".weight").dtype) ==
+            GGML_TYPE_BF16) {
+        return assets::TensorStorageType::F32;
+    }
+    return requested;
+}
+
 }  // namespace engine::models::minimax_music3

src/community_models/minimax_music3/condition_encoder.cpp

@@ -75,7 +75,7 @@ MiniMaxMusic3ConditionWeights load_condition_weights(
         *out.store,
         source,
         "proj",
-        storage_type,
+        conv_safe_storage_type(source, "proj", storage_type),
         config.out_dim,
         config.condition_hidden_dim,
         3,

src/community_models/minimax_music3/flow_transformer.cpp

@@ -56,7 +56,7 @@ MiniMaxMusic3FlowWeights load_flow_weights(
         *out.store,
         source,
         "preprocess_conv",
-        storage_type,
+        conv_safe_storage_type(source, "preprocess_conv", storage_type),
         concat_channels,
         concat_channels,
         1,
@@ -84,7 +84,7 @@ MiniMaxMusic3FlowWeights load_flow_weights(
         *out.store,
         source,
         "postprocess_conv",
-        storage_type,
+        conv_safe_storage_type(source, "postprocess_conv", storage_type),
         config.in_channels,
         config.in_channels,
         1,

src/community_models/minimax_music3/session.cpp

@@ -236,7 +236,7 @@ MiniMaxMusic3Request MiniMaxMusic3Session::parse_request(const runtime::TaskRequ
     if (out.lyrics.empty()) {
         throw std::runtime_error("MiniMax Music 3 requires lyrics");
     }
-    if (const auto value = runtime::parse_finite_float_option(request.options, {"duration_sec"})) {
+    if (const auto value = runtime::parse_finite_float_option(request.options, {"duration_sec", "duration_seconds"})) {
         out.duration_sec = *value;
     }
     if (const auto value = runtime::parse_i64_option(request.options, {"num_inference_steps"})) {

src/community_models/minimax_music3/vocoder.cpp

@@ -86,7 +86,7 @@ Music3VocoderWeights load_vocoder_weights(
         *out.store,
         source,
         "dec_in_proj",
-        storage_type,
+        conv_safe_storage_type(source, "dec_in_proj", storage_type),
         config.decoder_input_dim,
         config.latent_channels / 2,
         1,

webui/configs/model_params.json

@@ -123,6 +123,13 @@
     {"name": "timesignature", "type": "text", "label": "【曲谱】timesignature", "default": "", "placeholder": "如 4"}
   ],
 
+  "minimax_music3": [
+    {"name": "num_inference_steps", "type": "number", "label": "Flow steps per window", "default": 30, "minimum": 1, "maximum": 200, "step": 1, "precision": 0, "info": "Flow-matching Euler steps per 200-frame denoising window."},
+    {"name": "guidance_scale", "type": "slider", "label": "Flow guidance scale", "default": 1.7, "minimum": 0.0, "maximum": 10.0, "step": 0.1},
+    {"name": "ar_guidance_scale", "type": "slider", "label": "AR guidance scale", "default": 1.5, "minimum": 0.0, "maximum": 10.0, "step": 0.1, "info": "Classifier-free guidance of the semantic and residual code sampling."},
+    {"name": "top_k", "type": "number", "label": "top_k", "default": 50, "minimum": 1, "maximum": 1024, "step": 1, "precision": 0}
+  ],
+
   "minimax_h3": [
     {"name": "num_inference_steps", "type": "number", "label": "Denoising steps", "default": 12, "minimum": 1, "maximum": 50, "step": 1, "precision": 0, "info": "Twelve denoising steps provide a practical quality and performance balance."},
     {"name": "num_frames", "type": "number", "label": "Output frames", "default": 241, "minimum": 5, "maximum": 1441, "step": 4, "precision": 0, "info": "Approximately 24 frames per output second; 241 frames produces about 10 seconds of audio."},

webui/configs/models_catalog.json

@@ -56,6 +56,7 @@
     { "id": "chatterbox",             "display_name": "Chatterbox (voice clone)",              "family": "chatterbox",           "path": "models/chatterbox",                      "task": "clon",  "mode": "offline", "download_id": "chatterbox",              "min_vram_gb": 12 },
 
     { "id": "ace-step",               "display_name": "ACE-Step 1.5 (music gen)",              "family": "ace_step",             "path": "models/Ace-Step1.5",                     "task": "gen",   "mode": "offline", "download_id": "ace_step",                "session_options": { "ace_step.mem_saver": "true", "ace_step.dit_weight_type": "q8_0", "ace_step.text_encoder_weight_type": "q8_0", "ace_step.planner_weight_type": "q8_0" }, "min_vram_gb": 8 },
+    { "id": "minimax-music3",         "display_name": "MiniMax-Music3 (song gen)",             "family": "minimax_music3",       "path": "models/MiniMax-Music3-GGUF/language_model_q4_k.gguf", "task": "gen", "mode": "offline", "download_id": "minimax_music3_q4_k", "min_vram_gb": 12 },
     { "id": "stable-audio-small-music","display_name": "Stable Audio 3 Small Music (gen)",     "family": "stable_audio",         "path": "models/stable-audio-3-small-music",      "task": "gen",   "mode": "offline", "download_id": "stable_audio_3_small_music", "min_vram_gb": 4 },
     { "id": "stable-audio-small-sfx", "display_name": "Stable Audio 3 Small SFX (gen)",        "family": "stable_audio",         "path": "models/stable-audio-3-small-sfx",        "task": "gen",   "mode": "offline", "download_id": "stable_audio_3_small_sfx", "min_vram_gb": 4 },
     { "id": "stable-audio-medium",    "display_name": "Stable Audio 3 Medium (gen)",           "family": "stable_audio",         "path": "models/stable-audio-3-medium",           "task": "gen",   "mode": "offline", "download_id": "stable_audio_3_medium",   "session_options": { "stable_audio.mem_saver": "true" }, "min_vram_gb": 10 },

webui/native/src/lib/catalog.ts

@@ -118,6 +118,12 @@ function packageModelPath(entry: PackageEntry): string {
   if (entry.format === 'gguf' && entry.family === 'minimax_h3') {
     const entryName = entry.id.includes('int8_dit') ? 'dit_int8.gguf' : 'dit.gguf';
     modelFile = entry.files?.find((file) => file.toLowerCase().endsWith(`/${entryName}`));
+  } else if (entry.format === 'gguf' && entry.family === 'minimax_music3') {
+    // Multi-component package; the language_model_*.gguf matching the package
+    // precision is the entry file.
+    modelFile = entry.files?.find((file) =>
+      new RegExp(`/language_model_${entry.precision}\\.gguf$`).test(file.toLowerCase())) ||
+      entry.files?.find((file) => /\/language_model_[^/]*\.gguf$/.test(file.toLowerCase()));
   } else if (entry.format === 'gguf') {
     modelFile = entry.files?.find((file) => file.toLowerCase().endsWith('.gguf'));
   }