mtmd : Adding support for Nvidia Music Flamingo Model (#18470)

* Inital commit, debugging q5_k_s quant

* Made hf_to_gguf extend whisper to reduce code duplication

* addressed convert_hf_to_gguf pull request issue

---------

Co-authored-by: Henry D <henrydorsey147@gmail.com>

convert_hf_to_gguf.py

@@ -3503,7 +3503,7 @@ class QwenModel(TextModel):
         self._set_vocab_qwen()
 
 
-@ModelBase.register("Qwen2Model", "Qwen2ForCausalLM", "Qwen2AudioForConditionalGeneration", "KORMoForCausalLM")
+@ModelBase.register("Qwen2Model", "Qwen2ForCausalLM", "Qwen2AudioForConditionalGeneration", "KORMoForCausalLM", "AudioFlamingo3ForConditionalGeneration")
 class Qwen2Model(TextModel):
     model_arch = gguf.MODEL_ARCH.QWEN2
 
@@ -9292,6 +9292,18 @@ class VoxtralWhisperEncoderModel(WhisperEncoderModel):
         self.gguf_writer.add_audio_stack_factor(4) # == intermediate_size // hidden_size
 
 
+@ModelBase.register("AudioFlamingo3ForConditionalGeneration")
+class AudioFlamingo3WhisperEncoderModel(WhisperEncoderModel):
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        self.gguf_writer.add_clip_projector_type(gguf.VisionProjectorType.MUSIC_FLAMINGO)
+
+    def tensor_force_quant(self, name, new_name, bid, n_dims):
+        if ".conv" in name and ".weight" in name:
+            # Was trained in BF16, being safe, avoiding quantizing to FP16
+            return gguf.GGMLQuantizationType.F32
+        return super().tensor_force_quant(name, new_name, bid, n_dims)
+
 @ModelBase.register("FalconH1ForCausalLM")
 class FalconH1Model(Mamba2Model):
     model_arch = gguf.MODEL_ARCH.FALCON_H1

gguf-py/gguf/constants.py

@@ -3492,6 +3492,7 @@ class VisionProjectorType:
     COGVLM = "cogvlm"
     JANUS_PRO = "janus_pro"
     LFM2A = "lfm2a" # audio
+    MUSIC_FLAMINGO = "musicflamingo" # audio
     GLM4V = "glm4v"
 
 

tools/mtmd/clip-impl.h

@@ -180,6 +180,7 @@ enum projector_type {
     PROJECTOR_TYPE_GLMA,
     PROJECTOR_TYPE_QWEN25O, // will be replaced by QWEN2A or QWEN25VL depending on clip_ctx
     PROJECTOR_TYPE_VOXTRAL,
+    PROJECTOR_TYPE_MUSIC_FLAMINGO,
     PROJECTOR_TYPE_LFM2,
     PROJECTOR_TYPE_KIMIVL,
     PROJECTOR_TYPE_LIGHTONOCR,
@@ -209,6 +210,7 @@ static std::map<projector_type, std::string> PROJECTOR_TYPE_NAMES = {
     { PROJECTOR_TYPE_GLMA,      "glma"},
     { PROJECTOR_TYPE_QWEN25O,   "qwen2.5o"},
     { PROJECTOR_TYPE_VOXTRAL,   "voxtral"},
+    { PROJECTOR_TYPE_MUSIC_FLAMINGO, "musicflamingo"},
     { PROJECTOR_TYPE_LFM2,      "lfm2"},
     { PROJECTOR_TYPE_KIMIVL,    "kimivl"},
     { PROJECTOR_TYPE_LIGHTONOCR,"lightonocr"},

tools/mtmd/clip-model.h

@@ -319,7 +319,8 @@ struct clip_model {
 
     bool audio_has_avgpool() const {
         return proj_type == PROJECTOR_TYPE_QWEN2A
-            || proj_type == PROJECTOR_TYPE_VOXTRAL;
+            || proj_type == PROJECTOR_TYPE_VOXTRAL
+            || proj_type == PROJECTOR_TYPE_MUSIC_FLAMINGO;
     }
 
     bool audio_has_stack_frames() const {

tools/mtmd/clip.cpp

@@ -818,6 +818,7 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32
         case PROJECTOR_TYPE_VOXTRAL:
         case PROJECTOR_TYPE_QWEN2A:
         case PROJECTOR_TYPE_GLMA:
+        case PROJECTOR_TYPE_MUSIC_FLAMINGO:
             {
                 builder = std::make_unique<clip_graph_whisper_enc>(ctx, img);
             } break;
@@ -1176,6 +1177,7 @@ struct clip_model_loader {
                 case PROJECTOR_TYPE_QWEN2A:
                 case PROJECTOR_TYPE_GLMA:
                 case PROJECTOR_TYPE_VOXTRAL:
+                case PROJECTOR_TYPE_MUSIC_FLAMINGO:
                     {
                         bool require_stack = model.proj_type == PROJECTOR_TYPE_ULTRAVOX ||
                                              model.proj_type == PROJECTOR_TYPE_VOXTRAL ||
@@ -1576,6 +1578,17 @@ struct clip_model_loader {
                     model.mm_1_w = get_tensor(string_format(TN_MM_AUDIO_MLP, 1, "weight"));
                     model.mm_2_w = get_tensor(string_format(TN_MM_AUDIO_MLP, 2, "weight"));
                 } break;
+            case PROJECTOR_TYPE_MUSIC_FLAMINGO:
+                {
+                    model.conv1d_1_w = get_tensor(string_format(TN_CONV1D, 1, "weight"));
+                    model.conv1d_1_b = get_tensor(string_format(TN_CONV1D, 1, "bias"));
+                    model.conv1d_2_w = get_tensor(string_format(TN_CONV1D, 2, "weight"));
+                    model.conv1d_2_b = get_tensor(string_format(TN_CONV1D, 2, "bias"));
+                    model.mm_1_w = get_tensor(string_format(TN_MM_AUDIO_MLP, 1, "weight"));
+                    model.mm_1_b = get_tensor(string_format(TN_MM_AUDIO_MLP, 1, "bias"));
+                    model.mm_2_w = get_tensor(string_format(TN_MM_AUDIO_MLP, 2, "weight"));
+                    model.mm_2_b = get_tensor(string_format(TN_MM_AUDIO_MLP, 2, "bias"));
+                } break;
             case PROJECTOR_TYPE_INTERNVL:
                 {
                     model.mm_0_w = get_tensor(string_format(TN_MVLM_PROJ_MLP, 0, "weight"));
@@ -3031,6 +3044,7 @@ int clip_n_output_tokens(const struct clip_ctx * ctx, struct clip_image_f32 * im
         case PROJECTOR_TYPE_VOXTRAL:
         case PROJECTOR_TYPE_ULTRAVOX:
         case PROJECTOR_TYPE_QWEN2A:
+        case PROJECTOR_TYPE_MUSIC_FLAMINGO:
             {
                 n_patches = img->nx;
 
@@ -3403,6 +3417,7 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima
         case PROJECTOR_TYPE_ULTRAVOX:
         case PROJECTOR_TYPE_LFM2:
         case PROJECTOR_TYPE_VOXTRAL:
+        case PROJECTOR_TYPE_MUSIC_FLAMINGO:
         case PROJECTOR_TYPE_JANUS_PRO:
         case PROJECTOR_TYPE_COGVLM:
             {
@@ -3526,6 +3541,7 @@ int clip_n_mmproj_embd(const struct clip_ctx * ctx) {
             return ctx->model.projection->ne[1];
         case PROJECTOR_TYPE_ULTRAVOX:
         case PROJECTOR_TYPE_VOXTRAL:
+        case PROJECTOR_TYPE_MUSIC_FLAMINGO:
             return ctx->model.mm_2_w->ne[1];
         case PROJECTOR_TYPE_INTERNVL:
             return ctx->model.mm_3_w->ne[1];
@@ -3587,7 +3603,8 @@ bool clip_has_whisper_encoder(const struct clip_ctx * ctx) {
     return ctx->proj_type() == PROJECTOR_TYPE_ULTRAVOX
         || ctx->proj_type() == PROJECTOR_TYPE_QWEN2A
         || ctx->proj_type() == PROJECTOR_TYPE_GLMA
-        || ctx->proj_type() == PROJECTOR_TYPE_VOXTRAL;
+        || ctx->proj_type() == PROJECTOR_TYPE_VOXTRAL
+        || ctx->proj_type() == PROJECTOR_TYPE_MUSIC_FLAMINGO;
 }
 
 bool clip_encode_float_image (struct clip_ctx * ctx, int n_threads, float * img, int h, int w, float * vec) {

tools/mtmd/models/whisper-enc.cpp

@@ -86,6 +86,15 @@ ggml_cgraph * clip_graph_whisper_enc::build() {
             FFN_GELU_ERF,
             -1);
 
+    } else if (proj_type == PROJECTOR_TYPE_MUSIC_FLAMINGO) {
+        // projector
+        cur = build_ffn(cur,
+            model.mm_1_w, model.mm_1_b,
+            nullptr, nullptr,
+            model.mm_2_w, model.mm_2_b,
+            FFN_GELU_ERF,
+            -1);
+
     } else if (proj_type == PROJECTOR_TYPE_GLMA) {
             cur = ggml_norm(ctx0, cur, hparams.eps);
             cur = ggml_mul(ctx0, cur, model.mm_norm_pre_w);

tools/mtmd/mtmd.cpp

@@ -330,6 +330,7 @@ struct mtmd_context {
             case PROJECTOR_TYPE_ULTRAVOX:
             case PROJECTOR_TYPE_VOXTRAL:
             case PROJECTOR_TYPE_GLMA:
+            case PROJECTOR_TYPE_MUSIC_FLAMINGO:
                 audio_preproc = std::make_unique<mtmd_audio_preprocessor_whisper>(ctx_a);
                 break;
             case PROJECTOR_TYPE_LFM2A:
@@ -352,6 +353,9 @@ struct mtmd_context {
             // [BEGIN_AUDIO] ... (embeddings) ...
             aud_beg = "[BEGIN_AUDIO]";
 
+        } else if (proj == PROJECTOR_TYPE_MUSIC_FLAMINGO) {
+            // <sound> ... (embeddings) ...
+            aud_beg = "<sound>";
         }
     }