LLM inference in C/C++
C++ ggml
mtmd: support MiMo-V2.5 audio input (RVQ-based model) (#26190)
Xuan-Son Nguyen1cbfd1988311
diff --git a/conversion/mimo.py b/conversion/mimo.pyindex 11ec286794..ca2ed28ad3 100644--- a/conversion/mimo.py+++ b/conversion/mimo.py@@ -1,8 +1,9 @@from __future__ import annotations+import jsonimport re-from typing import Callable, TYPE_CHECKING+from typing import Any, Callable, Iterable, TYPE_CHECKINGimport torch@@ -229,7 +230,13 @@ class MimoV2Model(TextModel):@ModelBase.register("MiMoV2ForCausalLM")-class MiMoV2VisionModel(MmprojModel):+class MiMoV2VisionAudioModel(MmprojModel):+ has_audio_encoder = True++ _audio_tok_hparams: dict[str, Any] | None = None+ _rvq_codebook_sizes: list[int] | None = None+ _code_embd: dict[int, Tensor] | None = None+def __init__(self, *args, **kwargs):super().__init__(*args, **kwargs)assert self.hparams_vision is not None@@ -253,10 +260,22 @@ class MiMoV2VisionModel(MmprojModel):self.visual_token_window_size = int(hp.get("visual_token_window_size", -1))self.use_sink = bool(hp.get("use_sink", False))+ def get_audio_config(self) -> dict[str, Any] | None:+ if self._audio_tok_hparams is None:+ path = self.dir_model / "audio_tokenizer" / "config.json"+ with open(path, "r", encoding="utf-8") as f:+ cfg = json.load(f)+ # aliases so MmprojModel.find_aparam() / n_block_keys can resolve them+ cfg["hidden_size"] = cfg["d_model"]+ cfg["intermediate_size"] = cfg["encoder_ffn_dim"]+ cfg["num_attention_heads"] = cfg["encoder_attention_heads"]+ self._audio_tok_hparams = cfg+ return self._audio_tok_hparams+def set_gguf_parameters(self):super().set_gguf_parameters()- self.gguf_writer.add_clip_projector_type(gguf.VisionProjectorType.MIMOVL)+ self.gguf_writer.add_clip_vision_projector_type(gguf.VisionProjectorType.MIMOVL)self.gguf_writer.add_vision_use_silu(True)self.gguf_writer.add_vision_head_count_kv(self.num_kv_heads)self.gguf_writer.add_vision_spatial_merge_size(self.spatial_merge_size)@@ -266,19 +285,45 @@ class MiMoV2VisionModel(MmprojModel):self.gguf_writer.add_vision_min_pixels(int(self.preprocessor_config["min_pixels"]))self.gguf_writer.add_vision_max_pixels(int(self.preprocessor_config["max_pixels"]))+ assert self.hparams_audio is not None+ self.gguf_writer.add_clip_audio_projector_type(gguf.VisionProjectorType.MIMO_AUDIO)+ self.gguf_writer.add_audio_num_mel_bins(self.hparams_audio["n_mels"])+ self.gguf_writer.add_audio_attention_layernorm_eps(self.hparams_audio.get("layer_norm_eps", 1e-5))++ assert self._rvq_codebook_sizes is not None+ self.gguf_writer.add_audio_rvq_num_quantizers(len(self._rvq_codebook_sizes))+ self.gguf_writer.add_audio_rvq_codebook_size(self._rvq_codebook_sizes)++ n_layer = self.hparams_audio["encoder_layers"]+ swa_per_block = self.hparams_audio.get("swa_per_block", 1)+ if self.hparams_audio.get("hybrid_attention") and swa_per_block > 1:+ wa_pattern = [0 if i % swa_per_block < swa_per_block - 1 else -1 for i in range(n_layer)]+ else:+ wa_pattern = [-1] * n_layer+ self.gguf_writer.add_audio_wa_pattern_mode(wa_pattern)+ self.gguf_writer.add_audio_window_size(int(self.hparams_audio["encoder_attn_window_size"][0]))++ audio_cfg = self.global_config["audio_config"]+ self.gguf_writer.add_audio_local_block_count(int(audio_cfg["input_local_layers"]))+ self.gguf_writer.add_audio_local_group_size(int(audio_cfg["group_size"]))+def tensor_force_quant(self, name, new_name, bid, n_dims):- # Sinks must be F32: any sink-style softmax/mask add in ggml requires- # F32, and we fold sinks into a host-built F32 mask at encode time.- if new_name.endswith(".attn_sinks"):+ # for audio encoder: keep codebook in F32+ if new_name in (+ gguf.TENSOR_NAMES[gguf.MODEL_TENSOR.A_ENC_RVQ_CODEBOOK] + ".weight",+ gguf.TENSOR_NAMES[gguf.MODEL_TENSOR.A_MM_CODE_EMBD] + ".weight",+ ):+ return gguf.GGMLQuantizationType.F32+ if ("encoder.conv" in name or "encoder.down_sample_layer" in name) and name.endswith(".weight"):return gguf.GGMLQuantizationType.F32return super().tensor_force_quant(name, new_name, bid, n_dims)@classmethoddef filter_tensors(cls, item: tuple[str, Callable[[], Tensor]]) -> tuple[str, Callable[[], Tensor]] | None:name, _ = item- if not name.startswith("visual."):- return None- return super().filter_tensors(item)+ if name.startswith("visual.") or name.startswith("speech_embeddings.") or name.startswith("audio_encoder."):+ return super().filter_tensors(item)+ return Nonedef modify_tensors(self, data_torch, name, bid):# Conv3D patch embed: split along the temporal axis (kt=2) into two Conv2D@@ -292,4 +337,64 @@ class MiMoV2VisionModel(MmprojModel):yield (embd_name + ".weight.1", data_torch[:, :, 1, ...])return+ if m := re.match(r"^speech_embeddings\.(\d+)\.weight$", name):+ if self._code_embd is None:+ self._code_embd = {}+ self._code_embd[int(m.group(1))] = data_torch++ n_channels = int(self.global_config["audio_config"]["audio_channels"])+ if len(self._code_embd) < n_channels:+ return+ merged = torch.stack([self._code_embd.pop(i) for i in range(n_channels)], dim=0)+ yield (self.format_tensor_name(gguf.MODEL_TENSOR.A_MM_CODE_EMBD), merged)+ return++ if "conv1.bias" in name or "conv2.bias" in name:+ # transpose conv1/conv2 bias so it broadcasts against [n_frames, C_out, 1]+ data_torch = data_torch.unsqueeze(-1)++ if name == "audio_encoder.projection.mlp.0.weight":+ yield (self.format_tensor_name(gguf.MODEL_TENSOR.A_MMPROJ, 1), data_torch)+ return+ if name == "audio_encoder.projection.mlp.2.weight":+ yield (self.format_tensor_name(gguf.MODEL_TENSOR.A_MMPROJ, 2), data_torch)+ return+yield from super().modify_tensors(data_torch, name, bid)++ def generate_extra_tensors(self) -> Iterable[tuple[str, Tensor]]:+ # note: audio encoder is in its own subdir "audio_tokenizer"+ from safetensors.torch import load_file++ tok_dir = self.dir_model / "audio_tokenizer"+ state_dict = load_file(tok_dir / "model.safetensors")++ codebook_re = re.compile(r"^encoder\.quantizer\.vq\.layers\.(\d+)\._codebook\.embed$")+ codebooks: dict[int, Tensor] = {}++ # EMA/training-only RVQ buffers - not needed for inference (nearest-codebook+ # lookup only reads "_codebook.embed")+ skip_suffixes = (+ "_codebook.cluster_size",+ "_codebook.embed_avg",+ "_codebook.inited",+ )+ for name, tensor in state_dict.items():+ if name.endswith(skip_suffixes):+ continue+ if m := codebook_re.match(name):+ codebooks[int(m.group(1))] = tensor+ continue+ yield name, tensor++ # gather codebooks and merge into 3D tensor, similar to MoE MLP tensors+ n_q = len(codebooks)+ ordered = [codebooks[i] for i in range(n_q)]+ self._rvq_codebook_sizes = [int(cb.shape[0]) for cb in ordered]+ max_bins = max(self._rvq_codebook_sizes)+ dim = ordered[0].shape[1]+ merged = ordered[0].new_zeros(n_q, max_bins, dim)+ for i, cb in enumerate(ordered):+ merged[i, : cb.shape[0], :] = cb++ yield (self.format_tensor_name(gguf.MODEL_TENSOR.A_ENC_RVQ_CODEBOOK), merged)diff --git a/gguf-py/gguf/constants.py b/gguf-py/gguf/constants.pyindex 78e3c29a0d..f9264425f6 100644--- a/gguf-py/gguf/constants.py+++ b/gguf-py/gguf/constants.py@@ -376,6 +376,12 @@ class Keys:CONV_KERNEL_SIZE = "clip.audio.conv_kernel_size"MAX_POS_EMB = "clip.audio.max_pos_emb"FEATURE_LAYERS = "clip.audio.feature_layer" # Granite Speech Plus+ RVQ_NUM_QUANTIZERS = "clip.audio.rvq.num_quantizers"+ RVQ_CODEBOOK_SIZE = "clip.audio.rvq.codebook_size"+ WA_PATTERN_MODE = "clip.audio.wa_pattern_mode" # per-layer -1 (full) / 0 (windowed)+ WINDOW_SIZE = "clip.audio.window_size"+ LOCAL_BLOCK_COUNT = "clip.audio.local_block_count" # mimo-v2.5: input_local_transformer layer count+ LOCAL_GROUP_SIZE = "clip.audio.local_group_size" # mimo-v2.5: input_local_transformer grouping sizeclass Attention:HEAD_COUNT = "clip.audio.attention.head_count"@@ -945,6 +951,9 @@ class MODEL_TENSOR(IntEnum):A_ENC_FFN_SCALE_1 = auto() # gemma3nA_ENC_FFN_GATE_1 = auto() # lfm2, gemma3nA_ENC_FFN_DOWN_1 = auto() # lfm2, gemma3n+ A_ENC_DOWNSAMPLE_CONV = auto() # mimo-audio-tokenizer: post-transformer downsample conv+ A_ENC_DOWNSAMPLE_NORM = auto() # mimo-audio-tokenizer: post-transformer downsample norm+ A_ENC_RVQ_CODEBOOK = auto() # mimo-audio-tokenizer: residual vector quantizer codebook, per quantizer indexA_MMPROJ = auto()A_MMPROJ_FC = auto()A_MM_NORM_PRE = auto()@@ -953,6 +962,17 @@ class MODEL_TENSOR(IntEnum):A_MM_HARD_EMB_NORM = auto() # gemma3nA_MM_SOFT_EMB_NORM = auto() # gemma3nA_MM_INP_PROJ = auto() # gemma3n+ A_MM_CODE_EMBD = auto() # mimo: text-side RVQ code embedding table ("text codebook"), merged 3D [n_channels, vocab, dim]+ A_MM_LOCAL_ATTN_Q = auto() # mimo: input_local_transformer (LLM-side connector)+ A_MM_LOCAL_ATTN_K = auto()+ A_MM_LOCAL_ATTN_V = auto()+ A_MM_LOCAL_ATTN_OUT = auto()+ A_MM_LOCAL_FFN_GATE = auto()+ A_MM_LOCAL_FFN_UP = auto()+ A_MM_LOCAL_FFN_DOWN = auto()+ A_MM_LOCAL_LN1 = auto()+ A_MM_LOCAL_LN2 = auto()+ A_MM_LOCAL_NORM = auto() # final norm after all input_local_transformer layersA_PER_DIM_K_SCALE = auto() # gemma4A_PER_DIM_SCALE = auto() # gemma4# nextn/mtp@@ -1532,6 +1552,9 @@ TENSOR_NAMES: dict[MODEL_TENSOR, str] = {MODEL_TENSOR.A_ENC_FFN_UP_1: "a.blk.{bid}.ffn_up_1",MODEL_TENSOR.A_ENC_FFN_GATE_1: "a.blk.{bid}.ffn_gate_1",MODEL_TENSOR.A_ENC_FFN_DOWN_1: "a.blk.{bid}.ffn_down_1",+ MODEL_TENSOR.A_ENC_DOWNSAMPLE_CONV: "a.downsample.conv",+ MODEL_TENSOR.A_ENC_DOWNSAMPLE_NORM: "a.downsample.norm",+ MODEL_TENSOR.A_ENC_RVQ_CODEBOOK: "a.rvq.codebook",MODEL_TENSOR.A_MMPROJ: "mm.a.mlp.{bid}",MODEL_TENSOR.A_MMPROJ_FC: "mm.a.fc",MODEL_TENSOR.A_MM_NORM_PRE: "mm.a.norm_pre",@@ -1540,6 +1563,17 @@ TENSOR_NAMES: dict[MODEL_TENSOR, str] = {MODEL_TENSOR.A_MM_SOFT_EMB_NORM: "mm.a.soft_emb_norm", # gemma3nMODEL_TENSOR.A_MM_EMBEDDING: "mm.a.embedding", # gemma3nMODEL_TENSOR.A_MM_HARD_EMB_NORM: "mm.a.hard_emb_norm", # gemma3n+ MODEL_TENSOR.A_MM_CODE_EMBD: "mm.a.code_embd",+ MODEL_TENSOR.A_MM_LOCAL_ATTN_Q: "mm.a.local_blk.{bid}.attn_q",+ MODEL_TENSOR.A_MM_LOCAL_ATTN_K: "mm.a.local_blk.{bid}.attn_k",+ MODEL_TENSOR.A_MM_LOCAL_ATTN_V: "mm.a.local_blk.{bid}.attn_v",+ MODEL_TENSOR.A_MM_LOCAL_ATTN_OUT: "mm.a.local_blk.{bid}.attn_out",+ MODEL_TENSOR.A_MM_LOCAL_FFN_GATE: "mm.a.local_blk.{bid}.ffn_gate",+ MODEL_TENSOR.A_MM_LOCAL_FFN_UP: "mm.a.local_blk.{bid}.ffn_up",+ MODEL_TENSOR.A_MM_LOCAL_FFN_DOWN: "mm.a.local_blk.{bid}.ffn_down",+ MODEL_TENSOR.A_MM_LOCAL_LN1: "mm.a.local_blk.{bid}.ln1",+ MODEL_TENSOR.A_MM_LOCAL_LN2: "mm.a.local_blk.{bid}.ln2",+ MODEL_TENSOR.A_MM_LOCAL_NORM: "mm.a.local_norm",MODEL_TENSOR.A_PER_DIM_K_SCALE: "a.blk.{bid}.per_dim_k_scale", # gemma4MODEL_TENSOR.A_PER_DIM_SCALE: "a.blk.{bid}.per_dim_scale", # gemma4# lfm2 audio@@ -1741,10 +1775,24 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {MODEL_TENSOR.A_ENC_FFN_UP_1,MODEL_TENSOR.A_ENC_FFN_GATE_1,MODEL_TENSOR.A_ENC_FFN_DOWN_1,+ MODEL_TENSOR.A_ENC_DOWNSAMPLE_CONV,+ MODEL_TENSOR.A_ENC_DOWNSAMPLE_NORM,+ MODEL_TENSOR.A_ENC_RVQ_CODEBOOK,MODEL_TENSOR.A_MMPROJ,MODEL_TENSOR.A_MMPROJ_FC,MODEL_TENSOR.A_MM_NORM_PRE,MODEL_TENSOR.A_MM_NORM_MID,+ MODEL_TENSOR.A_MM_CODE_EMBD,+ MODEL_TENSOR.A_MM_LOCAL_ATTN_Q,+ MODEL_TENSOR.A_MM_LOCAL_ATTN_K,+ MODEL_TENSOR.A_MM_LOCAL_ATTN_V,+ MODEL_TENSOR.A_MM_LOCAL_ATTN_OUT,+ MODEL_TENSOR.A_MM_LOCAL_FFN_GATE,+ MODEL_TENSOR.A_MM_LOCAL_FFN_UP,+ MODEL_TENSOR.A_MM_LOCAL_FFN_DOWN,+ MODEL_TENSOR.A_MM_LOCAL_LN1,+ MODEL_TENSOR.A_MM_LOCAL_LN2,+ MODEL_TENSOR.A_MM_LOCAL_NORM,MODEL_TENSOR.A_ENC_NORM_CONV,MODEL_TENSOR.A_ENC_LINEAR_POS,MODEL_TENSOR.A_ENC_POS_BIAS_U,@@ -4804,6 +4852,7 @@ class VisionProjectorType:MINICPMV4_6 = "minicpmv4_6"GRANITE_SPEECH = "granite_speech" # audioMIMOVL = "mimovl"+ MIMO_AUDIO = "mimo_audio"GRANITE4_VISION = "granite4_vision"diff --git a/gguf-py/gguf/gguf_writer.py b/gguf-py/gguf/gguf_writer.pyindex ecf1f17ee8..bd8629aa11 100644--- a/gguf-py/gguf/gguf_writer.py+++ b/gguf-py/gguf/gguf_writer.py@@ -1350,6 +1350,24 @@ class GGUFWriter:def add_audio_num_mel_bins(self, value: int) -> None:self.add_uint32(Keys.ClipAudio.NUM_MEL_BINS, value)+ def add_audio_rvq_num_quantizers(self, value: int) -> None:+ self.add_uint32(Keys.ClipAudio.RVQ_NUM_QUANTIZERS, value)++ def add_audio_rvq_codebook_size(self, values: Sequence[int]) -> None:+ self.add_array(Keys.ClipAudio.RVQ_CODEBOOK_SIZE, values)++ def add_audio_wa_pattern_mode(self, modes: Sequence[int]) -> None:+ self.add_array(Keys.ClipAudio.WA_PATTERN_MODE, modes)++ def add_audio_window_size(self, value: int) -> None:+ self.add_uint32(Keys.ClipAudio.WINDOW_SIZE, value)++ def add_audio_local_block_count(self, value: int) -> None:+ self.add_uint32(Keys.ClipAudio.LOCAL_BLOCK_COUNT, value)++ def add_audio_local_group_size(self, value: int) -> None:+ self.add_uint32(Keys.ClipAudio.LOCAL_GROUP_SIZE, value)+def add_audio_stack_factor(self, value: int) -> None:self.add_uint32(Keys.ClipAudio.Projector.STACK_FACTOR, value)diff --git a/gguf-py/gguf/tensor_mapping.py b/gguf-py/gguf/tensor_mapping.pyindex 62d7a827e3..8299ac25b4 100644--- a/gguf-py/gguf/tensor_mapping.py+++ b/gguf-py/gguf/tensor_mapping.py@@ -2095,6 +2095,7 @@ class TensorNameMap:"conformer.pre_encode.conv.{bid}", # lfm2"model.audio_tower.subsample_conv_projection.conv_{bid}.conv", # gemma3n"conformer.subsample_conv_projection.layer{bid}.conv", # gemma4+ "encoder.conv{bid}", # mimo-audio-tokenizer),MODEL_TENSOR.A_ENC_CONV1D_NORM: (@@ -2119,6 +2120,7 @@ class TensorNameMap:MODEL_TENSOR.A_POST_NORM: ("audio_tower.layer_norm", # ultravox"audio_tower.ln_post", # qwen2omni+ "encoder.layer_norm", # mimo-audio-tokenizer),MODEL_TENSOR.A_ENC_ATTN_Q: (@@ -2127,6 +2129,7 @@ class TensorNameMap:"conformer.layers.{bid}.attention.attn.q_proj", # gemma3n"conformer.layers.{bid}.self_attn.q_proj", # gemma4"encoder.layers.{bid}.attn.to_q", # granite_speech+ "encoder.layers.{bid}.self_attn.q_proj", # mimo-audio-tokenizer),MODEL_TENSOR.A_ENC_ATTN_K: (@@ -2135,6 +2138,7 @@ class TensorNameMap:"conformer.layers.{bid}.attention.attn.k_proj", # gemma3n"conformer.layers.{bid}.self_attn.k_proj", # gemma4"encoder.layers.{bid}.attn.to_k", # granite_speech (split from to_kv)+ "encoder.layers.{bid}.self_attn.k_proj", # mimo-audio-tokenizer),MODEL_TENSOR.A_ENC_ATTN_V: (@@ -2143,6 +2147,7 @@ class TensorNameMap:"conformer.layers.{bid}.attention.attn.v_proj", # gemma3n"conformer.layers.{bid}.self_attn.v_proj", # gemma4"encoder.layers.{bid}.attn.to_v", # granite_speech (split from to_kv)+ "encoder.layers.{bid}.self_attn.v_proj", # mimo-audio-tokenizer),MODEL_TENSOR.A_ENC_ATTN_K_REL: (@@ -2171,6 +2176,7 @@ class TensorNameMap:"conformer.layers.{bid}.norm_self_att", # lfm2"conformer.layers.{bid}.attention.pre_attn_norm", # gemma3n"encoder.layers.{bid}.attn.pre_norm", # granite_speech+ "encoder.layers.{bid}.self_attn_layer_norm", # mimo-audio-tokenizer),MODEL_TENSOR.A_ENC_OUTPUT: (@@ -2179,6 +2185,7 @@ class TensorNameMap:"conformer.layers.{bid}.attention.post", # gemma3n"conformer.layers.{bid}.self_attn.post", # gemma4"encoder.layers.{bid}.attn.to_out", # granite_speech+ "encoder.layers.{bid}.self_attn.out_proj", # mimo-audio-tokenizer),MODEL_TENSOR.A_ENC_OUTPUT_NORM: (@@ -2186,6 +2193,7 @@ class TensorNameMap:"conformer.layers.{bid}.norm_out", # lfm2"conformer.layers.{bid}.attention.post_norm", # gemma3n"encoder.layers.{bid}.post_norm", # granite_speech+ "encoder.layers.{bid}.final_layer_norm", # mimo-audio-tokenizer),MODEL_TENSOR.A_ENC_FFN_NORM: (@@ -2210,6 +2218,7 @@ class TensorNameMap:"conformer.layers.{bid}.ffw_layer_start.ffw_layer_1", # gemma3n"conformer.layers.{bid}.feed_forward1.ffw_layer_1", # gemma4"encoder.layers.{bid}.ff1.up_proj", # granite_speech+ "encoder.layers.{bid}.fc1", # mimo-audio-tokenizer),MODEL_TENSOR.A_ENC_FFN_GATE: (),@@ -2220,6 +2229,7 @@ class TensorNameMap:"conformer.layers.{bid}.ffw_layer_start.ffw_layer_2", # gemma3n"conformer.layers.{bid}.feed_forward1.ffw_layer_2", # gemma4"encoder.layers.{bid}.ff1.down_proj", # granite_speech+ "encoder.layers.{bid}.fc2", # mimo-audio-tokenizer),MODEL_TENSOR.A_ENC_FFN_UP_1: (@@ -2243,6 +2253,19 @@ class TensorNameMap:"encoder.layers.{bid}.ff2.pre_norm", # granite_speech),+ MODEL_TENSOR.A_ENC_DOWNSAMPLE_CONV: (+ "encoder.down_sample_layer.0", # mimo-audio-tokenizer+ ),++ MODEL_TENSOR.A_ENC_DOWNSAMPLE_NORM: (+ "encoder.down_sample_norm", # mimo-audio-tokenizer+ ),++ # note: the raw per-quantizer "encoder.quantizer.vq.layers.{i}._codebook.embed"+ # tensors are merged (padded + stacked, like MoE experts) into this single 3D+ # tensor in conversion code, so no raw-name mapping is registered here.+ MODEL_TENSOR.A_ENC_RVQ_CODEBOOK: (),+MODEL_TENSOR.A_ENC_FFN_POST_NORM_1: ("conformer.layers.{bid}.ffw_layer_end.post_layer_norm", # gemma3n"conformer.layers.{bid}.feed_forward2.post_layer_norm", # gemma4@@ -2294,6 +2317,42 @@ class TensorNameMap:"audio.multi_modal_projector.ln_mid", # ultravox),+ # note: the raw per-channel "speech_embeddings.{i}" tensors are merged+ # (stacked, like MoE experts) into this single 3D tensor in conversion+ # code, so no raw-name mapping is registered here.+ MODEL_TENSOR.A_MM_CODE_EMBD: (),++ MODEL_TENSOR.A_MM_LOCAL_ATTN_Q: (+ "audio_encoder.input_local_transformer.layers.{bid}.self_attn.q_proj", # mimo-v2.5+ ),+ MODEL_TENSOR.A_MM_LOCAL_ATTN_K: (+ "audio_encoder.input_local_transformer.layers.{bid}.self_attn.k_proj", # mimo-v2.5+ ),+ MODEL_TENSOR.A_MM_LOCAL_ATTN_V: (+ "audio_encoder.input_local_transformer.layers.{bid}.self_attn.v_proj", # mimo-v2.5+ ),+ MODEL_TENSOR.A_MM_LOCAL_ATTN_OUT: (+ "audio_encoder.input_local_transformer.layers.{bid}.self_attn.o_proj", # mimo-v2.5+ ),+ MODEL_TENSOR.A_MM_LOCAL_FFN_GATE: (+ "audio_encoder.input_local_transformer.layers.{bid}.mlp.gate_proj", # mimo-v2.5+ ),+ MODEL_TENSOR.A_MM_LOCAL_FFN_UP: (+ "audio_encoder.input_local_transformer.layers.{bid}.mlp.up_proj", # mimo-v2.5+ ),+ MODEL_TENSOR.A_MM_LOCAL_FFN_DOWN: (+ "audio_encoder.input_local_transformer.layers.{bid}.mlp.down_proj", # mimo-v2.5+ ),+ MODEL_TENSOR.A_MM_LOCAL_LN1: (+ "audio_encoder.input_local_transformer.layers.{bid}.input_layernorm", # mimo-v2.5+ ),+ MODEL_TENSOR.A_MM_LOCAL_LN2: (+ "audio_encoder.input_local_transformer.layers.{bid}.post_attention_layernorm", # mimo-v2.5+ ),+ MODEL_TENSOR.A_MM_LOCAL_NORM: (+ "audio_encoder.input_local_transformer.norm", # mimo-v2.5+ ),+MODEL_TENSOR.A_ENC_CONV_DW: ("conformer.layers.{bid}.conv.depthwise_conv", # lfm2"conformer.layers.{bid}.lconv1d.depthwise_conv1d", # gemma3ndiff --git a/src/llama-quant.cpp b/src/llama-quant.cppindex 7c0bac07d0..92ebc11b99 100644--- a/src/llama-quant.cpp+++ b/src/llama-quant.cpp@@ -359,6 +359,10 @@ static bool tensor_allows_quantization(const llama_model_quantize_params * paramquantize &= name.find(".patch_embd") == std::string::npos;quantize &= name.find(".patch_merger") == std::string::npos;+ // audio codebook+ quantize &= name.find("a.rvq.codebook") == std::string::npos;+ quantize &= name.find("mm.a.code_embd") == std::string::npos;+return quantize;}diff --git a/tools/mtmd/CMakeLists.txt b/tools/mtmd/CMakeLists.txtindex fd7ddceb0b..18a8288ba0 100644--- a/tools/mtmd/CMakeLists.txt+++ b/tools/mtmd/CMakeLists.txt@@ -51,6 +51,7 @@ add_library(mtmdmodels/qwen3vl.cppmodels/mimovl.cppmodels/qwen3a.cpp+ models/mimo-audio.cppmodels/step3vl.cppmodels/siglip.cppmodels/whisper-enc.cppdiff --git a/tools/mtmd/clip-graph.h b/tools/mtmd/clip-graph.hindex a95de20a31..29352abb4c 100644--- a/tools/mtmd/clip-graph.h+++ b/tools/mtmd/clip-graph.h@@ -13,6 +13,14 @@struct build_vit_opts {ggml_tensor * attn_mask = nullptr;+ // TODO @ngxson : merge attn_mask and attn_mask_layers into one call+ std::vector<ggml_tensor *> attn_mask_layers; // one per layer++ // hook at layer output embeddings+ std::function<void(ggml_tensor * cur, int il)> callback_layer_out = nullptr;++ // whether to skip the automatic post-layernorm (model.post_ln_w) applied at the end+ bool skip_post_ln = false;};struct clip_graph {diff --git a/tools/mtmd/clip-impl.h b/tools/mtmd/clip-impl.hindex 42374311ce..0920411380 100644--- a/tools/mtmd/clip-impl.h+++ b/tools/mtmd/clip-impl.h@@ -82,6 +82,12 @@#define KEY_A_PROJ_WINDOW_SIZE "clip.audio.projector.window_size"#define KEY_A_PROJ_DOWNSAMPLE_RATE "clip.audio.projector.downsample_rate"#define KEY_A_PROJ_HEAD_COUNT "clip.audio.projector.head_count"+#define KEY_A_RVQ_NUM_QUANTIZERS "clip.audio.rvq.num_quantizers" // mimo-audio-tokenizer+#define KEY_A_RVQ_CODEBOOK_SIZE "clip.audio.rvq.codebook_size" // mimo-audio-tokenizer: per-quantizer bin count+#define KEY_A_WA_PATTERN_MODE "clip.audio.wa_pattern_mode" // mimo-audio-tokenizer, per-layer -1 (full) / 0 (windowed)+#define KEY_A_ATTN_WINDOW_SIZE "clip.audio.window_size" // mimo-audio-tokenizer: sliding-window radius+#define KEY_A_LOCAL_BLOCK_COUNT "clip.audio.local_block_count" // mimo-v2.5: input_local_transformer layer count+#define KEY_A_LOCAL_GROUP_SIZE "clip.audio.local_group_size" // mimo-v2.5: input_local_transformer grouping size//// tensor name constants@@ -175,6 +181,24 @@#define TN_MM_NORM_PRE "mm.a.norm_pre.%s"#define TN_MM_NORM_MID "mm.a.norm_mid.%s"+// mimo-audio-tokenizer+#define TN_A_DOWNSAMPLE_CONV "a.downsample.conv.%s"+#define TN_A_DOWNSAMPLE_NORM "a.downsample.norm.%s"+#define TN_A_RVQ_CODEBOOK "a.rvq.codebook.%s"+// mimo-v2.5: text-side RVQ code embedding ("text codebook")+#define TN_MM_A_CODE_EMBD "mm.a.code_embd.%s"+// mimo-v2.5: LLM-side connector (input_local_transformer)+#define TN_MM_A_LOCAL_ATTN_Q "mm.a.local_blk.%d.attn_q.%s"+#define TN_MM_A_LOCAL_ATTN_K "mm.a.local_blk.%d.attn_k.%s"+#define TN_MM_A_LOCAL_ATTN_V "mm.a.local_blk.%d.attn_v.%s"+#define TN_MM_A_LOCAL_ATTN_OUT "mm.a.local_blk.%d.attn_out.%s"+#define TN_MM_A_LOCAL_FFN_GATE "mm.a.local_blk.%d.ffn_gate.%s"+#define TN_MM_A_LOCAL_FFN_UP "mm.a.local_blk.%d.ffn_up.%s"+#define TN_MM_A_LOCAL_FFN_DOWN "mm.a.local_blk.%d.ffn_down.%s"+#define TN_MM_A_LOCAL_LN1 "mm.a.local_blk.%d.ln1.%s"+#define TN_MM_A_LOCAL_LN2 "mm.a.local_blk.%d.ln2.%s"+#define TN_MM_A_LOCAL_NORM "mm.a.local_norm.%s"+// cogvlm#define TN_MM_POST_FC_NORM "mm.post_fc_norm.%s"#define TN_MM_H_TO_4H "mm.up.%s"@@ -374,6 +398,7 @@ enum projector_type {PROJECTOR_TYPE_MIMOVL,PROJECTOR_TYPE_MINIMAX_M3,PROJECTOR_TYPE_GRANITE4_VISION,+ PROJECTOR_TYPE_MIMO_AUDIO,PROJECTOR_TYPE_UNKNOWN,};@@ -429,6 +454,7 @@ static std::map<projector_type, std::string> PROJECTOR_TYPE_NAMES = {{ PROJECTOR_TYPE_MIMOVL, "mimovl"},{ PROJECTOR_TYPE_MINIMAX_M3, "minimax_m3"},{ PROJECTOR_TYPE_GRANITE4_VISION, "granite4_vision"},+ { PROJECTOR_TYPE_MIMO_AUDIO, "mimo_audio"},};static projector_type clip_projector_type_from_string(const std::string & str) {diff --git a/tools/mtmd/clip-model.h b/tools/mtmd/clip-model.hindex 850957d7de..8dc8754976 100644--- a/tools/mtmd/clip-model.h+++ b/tools/mtmd/clip-model.h@@ -124,6 +124,14 @@ struct clip_hparams {int32_t audio_window_len = -1;int32_t audio_hop_len = -1;+ // mimo-audio-tokenizer: residual vector quantizer+ int32_t rvq_num_quantizers = 0;+ std::vector<int32_t> rvq_codebook_size; // per-quantizer bin count (ragged, e.g. 1024/1024/256/128x17)++ // mimo-v2.5: LLM-side connector (input_local_transformer)+ int32_t audio_local_n_layer = 0;+ int32_t audio_local_group_size = 0;+// legacybool has_llava_projector = false;int minicpmv_version = 0;@@ -537,6 +545,20 @@ struct clip_model {ggml_tensor * mm_norm_pre_b = nullptr;ggml_tensor * mm_norm_mid_w = nullptr;+ // mimo-audio-tokenizer: post-transformer downsample + RVQ codebook+ ggml_tensor * downsample_conv_w = nullptr; // no bias+ ggml_tensor * downsample_norm_w = nullptr;+ ggml_tensor * downsample_norm_b = nullptr;+ ggml_tensor * rvq_codebook = nullptr; // merged 3D [n_q, max_bins, dim]++ // mimo-v2.5: text-side RVQ code embedding ("text codebook")+ ggml_tensor * mm_a_code_embd = nullptr; // merged 3D [n_channels, vocab, dim]++ // mimo-v2.5: LLM-side connector (input_local_transformer, separate from the+ // audio_tokenizer's own encoder `layers`)+ std::vector<clip_layer> mm_a_local_layers;+ ggml_tensor * mm_a_local_norm_w = nullptr;+// qwen3aggml_tensor * conv2d_1_w = nullptr;ggml_tensor * conv2d_1_b = nullptr;diff --git a/tools/mtmd/clip.cpp b/tools/mtmd/clip.cppindex e0e2107a0b..04614b93bd 100644--- a/tools/mtmd/clip.cpp+++ b/tools/mtmd/clip.cpp@@ -340,6 +340,11 @@ ggml_tensor * clip_graph::build_vit(auto & layer = model.layers[il];ggml_tensor * cur = inpL; // inpL = residual, cur = hidden_states+ ggml_tensor * attn_mask = opts.attn_mask;+ if (opts.attn_mask_layers.size() > (size_t) il) {+ attn_mask = opts.attn_mask_layers[il];+ }+// layernorm1cur = build_norm(cur, layer.ln_1_w, layer.ln_1_b, norm_t, eps, il);cb(cur, "layer_inp_normed", il);@@ -452,7 +457,7 @@ ggml_tensor * clip_graph::build_vit(// build_attn returns a flat 2D [n_embd, n_pos*B]cur = build_attn(layer.o_w, layer.o_b,- Qcur, Kcur, Vcur, opts.attn_mask, kq_scale, il);+ Qcur, Kcur, Vcur, attn_mask, kq_scale, il);cb(cur, "attn_out", il);}@@ -471,6 +476,10 @@ ggml_tensor * clip_graph::build_vit(inpL = cur; // inpL = residual, cur = hidden_states+ if (opts.callback_layer_out) {+ opts.callback_layer_out(cur, il);+ }+cb(cur, "ffn_inp", il);// layernorm2 (pre-ffn norm)@@ -519,7 +528,7 @@ ggml_tensor * clip_graph::build_vit(}// post-layernorm- if (model.post_ln_w) {+ if (model.post_ln_w && !opts.skip_post_ln) {inpL = build_norm(inpL, model.post_ln_w, model.post_ln_b, norm_t, eps, -1);}@@ -1012,6 +1021,10 @@ static std::unique_ptr<clip_graph> clip_get_graph_builder(clip_ctx * ctx, const{builder = std::make_unique<clip_graph_qwen3a>(ctx, img);} break;+ case PROJECTOR_TYPE_MIMO_AUDIO:+ {+ builder = std::make_unique<clip_graph_mimo_audio>(ctx, img);+ } break;case PROJECTOR_TYPE_YOUTUVL:{builder = std::make_unique<clip_graph_youtuvl>(ctx, img);@@ -1575,6 +1588,45 @@ struct clip_model_loader {hparams.audio_window_len = 400;hparams.audio_hop_len = 160;} break;+ case PROJECTOR_TYPE_MIMO_AUDIO:+ {+ get_u32(KEY_A_RVQ_NUM_QUANTIZERS, hparams.rvq_num_quantizers, false);+ get_arr_int(KEY_A_RVQ_CODEBOOK_SIZE, hparams.rvq_codebook_size, false);+ if (hparams.rvq_num_quantizers <= 0) {+ throw std::runtime_error(string_format("%s: mimo_audio: missing %s\n", __func__, KEY_A_RVQ_NUM_QUANTIZERS));+ }+ if ((int) hparams.rvq_codebook_size.size() != hparams.rvq_num_quantizers) {+ throw std::runtime_error(string_format(+ "%s: mimo_audio: %s length (%zu) must equal %s (%d)\n", __func__,+ KEY_A_RVQ_CODEBOOK_SIZE, hparams.rvq_codebook_size.size(),+ KEY_A_RVQ_NUM_QUANTIZERS, hparams.rvq_num_quantizers));+ }+ hparams.ffn_op = FFN_GELU_ERF; // PyTorch F.gelu default (approximate="none")+ hparams.rope_theta = 10000.0f;++ // audio preprocessing params (mel spectrogram)+ hparams.audio_sample_rate = 24000;+ hparams.audio_n_fft = 960;+ hparams.audio_window_len = 960;+ hparams.audio_hop_len = 240;++ get_u32(KEY_A_ATTN_WINDOW_SIZE, hparams.attn_window_size);+ std::vector<int> wa_pattern;+ get_arr_int(KEY_A_WA_PATTERN_MODE, wa_pattern, true);+ if ((int) wa_pattern.size() != hparams.n_layer) {+ throw std::runtime_error(string_format(+ "%s: mimo_audio: %s length (%zu) must equal n_layer (%d)\n", __func__,+ KEY_A_WA_PATTERN_MODE, wa_pattern.size(), hparams.n_layer));+ }+ hparams.wa_pattern_mode.assign(wa_pattern.begin(), wa_pattern.end());++ get_u32(KEY_A_LOCAL_BLOCK_COUNT, hparams.audio_local_n_layer);+ get_u32(KEY_A_LOCAL_GROUP_SIZE, hparams.audio_local_group_size);+ if (hparams.audio_local_group_size <= 0) {+ throw std::runtime_error(string_format(+ "%s: mimo_audio: %s must be > 0\n", __func__, KEY_A_LOCAL_GROUP_SIZE));+ }+ } break;case PROJECTOR_TYPE_PADDLEOCR:{hparams.n_merge = 2;@@ -2444,6 +2496,54 @@ struct clip_model_loader {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_MIMO_AUDIO:+ {+ 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.downsample_conv_w = get_tensor(string_format(TN_A_DOWNSAMPLE_CONV, "weight"));+ model.downsample_norm_w = get_tensor(string_format(TN_A_DOWNSAMPLE_NORM, "weight"));+ model.downsample_norm_b = get_tensor(string_format(TN_A_DOWNSAMPLE_NORM, "bias"));+ model.rvq_codebook = get_tensor(string_format(TN_A_RVQ_CODEBOOK, "weight"), false);+ model.mm_a_code_embd = get_tensor(string_format(TN_MM_A_CODE_EMBD, "weight"), false);+ if (!model.rvq_codebook || !model.mm_a_code_embd) {+ throw std::runtime_error(string_format("%s: mimo_audio: missing %s or %s\n", __func__,+ TN_A_RVQ_CODEBOOK, TN_MM_A_CODE_EMBD));+ }+ // hparams.rvq_codebook_size comes from GGUF metadata and is independent of the+ // tensors' actual shapes - bound it so codebook/code_embd views built from it+ // (mimo-audio.cpp) can never read past either tensor's allocated bins/vocab.+ for (int32_t bins : hparams.rvq_codebook_size) {+ if (bins <= 0 || bins > model.rvq_codebook->ne[1] || bins > model.mm_a_code_embd->ne[1]) {+ throw std::runtime_error(string_format(+ "%s: mimo_audio: %s entry (%d) out of range for codebook/code_embd tensors\n",+ __func__, KEY_A_RVQ_CODEBOOK_SIZE, bins));+ }+ }++ // LLM-side connector: input_local_transformer + projection+ model.mm_a_local_layers.resize(hparams.audio_local_n_layer);+ for (int il = 0; il < hparams.audio_local_n_layer; il++) {+ auto & layer = model.mm_a_local_layers[il];+ layer.q_w = get_tensor(string_format(TN_MM_A_LOCAL_ATTN_Q, il, "weight"));+ layer.q_b = get_tensor(string_format(TN_MM_A_LOCAL_ATTN_Q, il, "bias"));+ layer.k_w = get_tensor(string_format(TN_MM_A_LOCAL_ATTN_K, il, "weight"));+ layer.k_b = get_tensor(string_format(TN_MM_A_LOCAL_ATTN_K, il, "bias"));+ layer.v_w = get_tensor(string_format(TN_MM_A_LOCAL_ATTN_V, il, "weight"));+ layer.v_b = get_tensor(string_format(TN_MM_A_LOCAL_ATTN_V, il, "bias"));+ layer.o_w = get_tensor(string_format(TN_MM_A_LOCAL_ATTN_OUT, il, "weight"));+ layer.ff_gate_w = get_tensor(string_format(TN_MM_A_LOCAL_FFN_GATE, il, "weight"));+ layer.ff_up_w = get_tensor(string_format(TN_MM_A_LOCAL_FFN_UP, il, "weight"));+ layer.ff_down_w = get_tensor(string_format(TN_MM_A_LOCAL_FFN_DOWN, il, "weight"));+ layer.ln_1_w = get_tensor(string_format(TN_MM_A_LOCAL_LN1, il, "weight"));+ layer.ln_2_w = get_tensor(string_format(TN_MM_A_LOCAL_LN2, il, "weight"));+ }+ model.mm_a_local_norm_w = get_tensor(string_format(TN_MM_A_LOCAL_NORM, "weight"));++ 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_VOXTRAL:{model.conv1d_1_w = get_tensor(string_format(TN_CONV1D, 1, "weight"));@@ -3549,6 +3649,15 @@ int clip_n_output_tokens(const clip_ctx * ctx, const clip_image_f32 * img) {{n_patches = img->nx(); // no downsampling: one token per raw waveform frame} break;+ case PROJECTOR_TYPE_MIMO_AUDIO:+ {+ // conv1(s=1) + conv2(s=2) -> RVQ-encoder downsample conv(k=2,s=2)+ int n = img->nx();+ n = (n - 1) / 2 + 1; // conv1 + conv2+ n = (n - 2) / 2 + 1; // downsample conv+ const int group_size = params.audio_local_group_size;+ n_patches = (n + group_size - 1) / group_size;+ } break;case PROJECTOR_TYPE_GRANITE_SPEECH:{const int ws = ctx->model.hparams.audio_proj_window_size;@@ -4376,6 +4485,58 @@ bool clip_image_batch_encode(clip_ctx * ctx, int n_threads, const clip_image_f32set_input_f32("pos_emb", pos_emb);}} break;+ case PROJECTOR_TYPE_MIMO_AUDIO:+ {+ GGML_ASSERT(imgs.entries.size() == 1);+ const int n_frames = imgs.entries.front().nx();+ const int n_pos = (n_frames - 1) / 2 + 1; // matches conv1(s=1)+conv2(s=2) output length++ std::vector<int32_t> positions(n_pos);+ for (int i = 0; i < n_pos; i++) {+ positions[i] = i;+ }+ set_input_i32("mimo_audio_positions", positions);++ const int window = hparams.attn_window_size;+ GGML_ASSERT(window > 0);++ const float neg_inf = std::numeric_limits<float>::lowest();+ std::vector<float> full_mask((size_t) n_pos * n_pos);+ std::vector<float> window_mask((size_t) n_pos * n_pos);+ for (int q = 0; q < n_pos; q++) {+ for (int k = 0; k < n_pos; k++) {+ const bool causal_ok = k <= q;+ full_mask[(size_t) q * n_pos + k] = causal_ok ? 0.0f : neg_inf;+ window_mask[(size_t) q * n_pos + k] = (causal_ok && (q - k) <= window) ? 0.0f : neg_inf;+ }+ }+ set_input_f32("mimo_audio_full_mask", full_mask);+ set_input_f32("mimo_audio_window_mask", window_mask);++ // input_local_transformer: block-diagonal mask + in-group positions+ {+ const int n_pos_ds = (n_pos - 2) / 2 + 1; // matches downsample conv (k=2,s=2,p=0)+ const int group_size = hparams.audio_local_group_size;+ GGML_ASSERT(group_size > 0);+ const int n_groups = (n_pos_ds + group_size - 1) / group_size;+ const int n_padded = n_groups * group_size;++ std::vector<int32_t> local_positions(n_padded);+ for (int i = 0; i < n_padded; i++) {+ local_positions[i] = i % group_size;+ }+ set_input_i32("mimo_audio_local_positions", local_positions);++ std::vector<float> local_mask((size_t) n_padded * n_padded);+ for (int q = 0; q < n_padded; q++) {+ for (int k = 0; k < n_padded; k++) {+ const bool same_group = (q / group_size) == (k / group_size);+ local_mask[(size_t) q * n_padded + k] = same_group ? 0.0f : neg_inf;+ }+ }+ set_input_f32("mimo_audio_local_mask", local_mask);+ }+ } break;case PROJECTOR_TYPE_LFM2A:{GGML_ASSERT(imgs.entries.size() == 1);@@ -4678,6 +4839,8 @@ int clip_n_mmproj_embd(const struct clip_ctx * ctx) {return ctx->model.qf_proj_blocks.size() * ctx->model.hparams.projection_dim;case PROJECTOR_TYPE_GLM4V:return ctx->model.mm_ffn_down_w->ne[1];+ case PROJECTOR_TYPE_MIMO_AUDIO:+ return ctx->model.mm_2_w->ne[1];default:GGML_ABORT("Unknown projector type");}diff --git a/tools/mtmd/models/mimo-audio.cpp b/tools/mtmd/models/mimo-audio.cppnew file mode 100644index 0000000000..481b36cc8d--- /dev/null+++ b/tools/mtmd/models/mimo-audio.cpp@@ -0,0 +1,218 @@+#include "models.h"++ggml_cgraph * clip_graph_mimo_audio::build() {+ ggml_tensor * inp = build_inp_raw(1); // [n_frames, n_mel, 1]++ ggml_tensor * cur = ggml_conv_1d_ph(ctx0, model.conv1d_1_w, inp, 1, 1);+ cur = ggml_add(ctx0, cur, model.conv1d_1_b);+ cur = ggml_gelu_erf(ctx0, cur);++ cur = ggml_conv_1d_ph(ctx0, model.conv1d_2_w, cur, 2, 1);+ cur = ggml_add(ctx0, cur, model.conv1d_2_b);+ cur = ggml_gelu_erf(ctx0, cur);++ ggml_tensor * inpL = ggml_cont(ctx0, ggml_transpose(ctx0, cur)); // [n_embd, n_pos]+ const int64_t n_pos = inpL->ne[1];+ cb(inpL, "after_conv1d", -1);++ GGML_ASSERT((int) hparams.wa_pattern_mode.size() == n_layer);++ ggml_tensor * inp_pos = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_pos);+ ggml_set_name(inp_pos, "mimo_audio_positions");+ ggml_set_input(inp_pos);++ ggml_tensor * full_mask = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_pos, n_pos);+ ggml_set_name(full_mask, "mimo_audio_full_mask");+ ggml_set_input(full_mask);++ ggml_tensor * window_mask = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_pos, n_pos);+ ggml_set_name(window_mask, "mimo_audio_window_mask");+ ggml_set_input(window_mask);++ build_vit_opts opts;+ opts.attn_mask_layers.resize(n_layer);+ for (int il = 0; il < n_layer; il++) {+ opts.attn_mask_layers[il] = hparams.wa_pattern_mode[il] == -1 ? full_mask : window_mask;+ }+ // the skip connection below must be added before the post-transformer norm,+ // so build_vit must not apply that norm itself+ opts.skip_post_ln = true;++ // encoder_skip_layer_id=3 (1-indexed) -> capture output of layer index 2+ const int skip_capture_il = 2;+ GGML_ASSERT(n_layer > skip_capture_il);+ ggml_tensor * skip_hidden = nullptr;+ opts.callback_layer_out = [&](ggml_tensor * layer_cur, int il) {+ if (il == skip_capture_il) {+ skip_hidden = layer_cur;+ }+ };++ auto add_pos = [&](ggml_tensor * x, const clip_layer &) {+ return ggml_rope_ext(ctx0, x, inp_pos, nullptr, d_head,+ GGML_ROPE_TYPE_NEOX, 0, hparams.rope_theta, 1.0f, 0.0f, 1.0f, 0.0f, 0.0f);+ };++ inpL = build_vit(inpL, n_pos, NORM_TYPE_NORMAL, hparams.ffn_op, nullptr, add_pos, opts);+ inpL = ggml_reshape_2d(ctx0, inpL, n_embd, n_pos); // build_vit restores a (size-1) batch dim++ GGML_ASSERT(skip_hidden != nullptr);+ inpL = ggml_add(ctx0, inpL, skip_hidden);++ inpL = build_norm(inpL, model.post_ln_w, model.post_ln_b, NORM_TYPE_NORMAL, eps, -1);+ cb(inpL, "after_transformer", -1);++ // downsample: strided conv (no bias) + gelu + layernorm+ {+ ggml_tensor * ds = ggml_cont(ctx0, ggml_transpose(ctx0, inpL)); // [n_pos, n_embd]+ ds = ggml_conv_1d(ctx0, model.downsample_conv_w, ds, 2, 0, 1);+ ds = ggml_gelu_erf(ctx0, ds);+ ds = ggml_cont(ctx0, ggml_transpose(ctx0, ds)); // [n_embd, n_pos/2]+ ds = build_norm(ds, model.downsample_norm_w, model.downsample_norm_b, NORM_TYPE_NORMAL, eps, -1);+ inpL = ds;+ }+ cb(inpL, "after_downsample", -1);++ // RVQ quantize: codebook ne=[dim, max_bins, n_q]+ // quantize input vector to codes (type=I32)+ std::vector<ggml_tensor *> codes;+ {+ GGML_ASSERT(model.rvq_codebook != nullptr);+ const int64_t dim = model.rvq_codebook->ne[0];+ GGML_ASSERT(dim == inpL->ne[0]);+ GGML_ASSERT((int64_t) hparams.rvq_codebook_size.size() == model.rvq_codebook->ne[2]);++ ggml_tensor * residual = inpL; // [dim, n_pos_ds]++ for (size_t q = 0; q < hparams.rvq_codebook_size.size(); q++) {+ const int64_t bins = hparams.rvq_codebook_size[q];+ ggml_tensor * codebook_q = ggml_view_2d(ctx0, model.rvq_codebook, dim, bins,+ model.rvq_codebook->nb[1], q * model.rvq_codebook->nb[2]);+ codebook_q = ggml_cont(ctx0, codebook_q);++ ggml_tensor * codebook_norm = ggml_sum_rows(ctx0, ggml_sqr(ctx0, codebook_q)); // [1, bins]+ codebook_norm = ggml_cont(ctx0, ggml_transpose(ctx0, codebook_norm)); // [bins, 1]++ ggml_tensor * dot = ggml_mul_mat(ctx0, codebook_q, residual); // [bins, n_pos_ds]+ ggml_tensor * scores = ggml_sub(ctx0, ggml_scale(ctx0, dot, 2.0f), codebook_norm);++ ggml_tensor * idx = ggml_argmax(ctx0, scores); // [n_pos_ds]+ codes.push_back(idx);++ ggml_tensor * quant = ggml_get_rows(ctx0, codebook_q, idx); // [dim, n_pos_ds]+ residual = ggml_sub(ctx0, residual, quant);+ cb(idx, "rvq_code", (int) q);+ }+ }++ // convert codes to LLM embeddings+ ggml_tensor * code_embd_sum = nullptr;+ {+ GGML_ASSERT(model.mm_a_code_embd != nullptr);+ const int64_t dim = model.mm_a_code_embd->ne[0];+ const int64_t vocab = model.mm_a_code_embd->ne[1];+ GGML_ASSERT((int64_t) codes.size() == model.mm_a_code_embd->ne[2]);+ GGML_ASSERT(dim == inpL->ne[0]);++ for (size_t i = 0; i < codes.size(); i++) {+ ggml_tensor * table_i = ggml_view_2d(ctx0, model.mm_a_code_embd, dim, vocab,+ model.mm_a_code_embd->nb[1], i * model.mm_a_code_embd->nb[2]);+ table_i = ggml_cont(ctx0, table_i);++ ggml_tensor * embd_i = ggml_get_rows(ctx0, table_i, codes[i]); // [dim, n_pos_ds]+ code_embd_sum = code_embd_sum ? ggml_add(ctx0, code_embd_sum, embd_i) : embd_i;+ }+ cb(code_embd_sum, "code_embd_sum", -1);+ }++ // input_local_transformer+ // groups of `group_size` consecutive downsampled frames are processed together, attending only within their own group.+ // Implemented as a block-diagonal mask + in-group-repeating positions+ // (rather than a real batch dim) - same technique as the encoder's masks above, and as gemma4a's / deepseekocr2's chunked attention.++ // note: hand-rolled here instead of build_vit() because this is a second, independent layer stack+ // (own layer array/count, RMSNorm instead of LN, SiLU FFN, own RoPE theta)++ ggml_tensor * projected;+ {+ const int group_size = hparams.audio_local_group_size;+ GGML_ASSERT(group_size > 0);+ const int64_t n_pos_ds = code_embd_sum->ne[1];+ const int64_t n_groups = (n_pos_ds + group_size - 1) / group_size;+ const int64_t n_padded = n_groups * group_size;++ ggml_tensor * cur_local = code_embd_sum;+ if (n_padded != n_pos_ds) {+ cur_local = ggml_pad(ctx0, cur_local, 0, (int) (n_padded - n_pos_ds), 0, 0);+ }++ ggml_tensor * local_pos = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_padded);+ ggml_set_name(local_pos, "mimo_audio_local_positions");+ ggml_set_input(local_pos);++ ggml_tensor * local_mask = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_padded, n_padded);+ ggml_set_name(local_mask, "mimo_audio_local_mask");+ ggml_set_input(local_mask);++ const float local_rope_theta = 640000.0f; // audio_config.rope_theta (differs from the encoder's)+ auto apply_local_rope = [&](ggml_tensor * x) {+ return ggml_rope_ext(ctx0, x, local_pos, nullptr, d_head,+ GGML_ROPE_TYPE_NEOX, 0, local_rope_theta, 1.0f, 0.0f, 1.0f, 0.0f, 0.0f);+ };++ for (int il = 0; il < hparams.audio_local_n_layer; il++) {+ auto & layer = model.mm_a_local_layers[il];++ ggml_tensor * attn_in = build_norm(cur_local, layer.ln_1_w, nullptr, NORM_TYPE_RMS, eps, il);++ ggml_tensor * Qcur = build_mm(layer.q_w, attn_in);+ if (layer.q_b) {+ Qcur = ggml_add(ctx0, Qcur, layer.q_b);+ }+ ggml_tensor * Kcur = build_mm(layer.k_w, attn_in);+ if (layer.k_b) {+ Kcur = ggml_add(ctx0, Kcur, layer.k_b);+ }+ ggml_tensor * Vcur = build_mm(layer.v_w, attn_in);+ if (layer.v_b) {+ Vcur = ggml_add(ctx0, Vcur, layer.v_b);+ }++ Qcur = ggml_reshape_3d(ctx0, Qcur, d_head, n_head, n_padded);+ Kcur = ggml_reshape_3d(ctx0, Kcur, d_head, n_head, n_padded);+ Vcur = ggml_reshape_3d(ctx0, Vcur, d_head, n_head, n_padded);++ Qcur = apply_local_rope(Qcur);+ Kcur = apply_local_rope(Kcur);++ ggml_tensor * attn_out = build_attn(layer.o_w, nullptr, Qcur, Kcur, Vcur, local_mask, kq_scale, il);+ cur_local = ggml_add(ctx0, cur_local, attn_out);++ ggml_tensor * ffn_in = build_norm(cur_local, layer.ln_2_w, nullptr, NORM_TYPE_RMS, eps, il);+ ggml_tensor * ffn_out = build_ffn(ffn_in,+ layer.ff_up_w, nullptr,+ layer.ff_gate_w, nullptr,+ layer.ff_down_w, nullptr,+ FFN_SILU, il);+ cur_local = ggml_add(ctx0, cur_local, ffn_out);+ }++ cur_local = build_norm(cur_local, model.mm_a_local_norm_w, nullptr, NORM_TYPE_RMS, eps, -1);+ cb(cur_local, "after_local_transformer", -1);++ // flatten each group of `group_size` frames into one (group_size*n_embd)-dim vector+ // (matching AudioProjection's flattened input)+ ggml_tensor * grouped = ggml_reshape_2d(ctx0, cur_local, n_embd * group_size, n_groups);++ // AudioProjection: Linear (no bias) -> GELU -> Linear (no bias)+ projected = build_ffn(grouped,+ model.mm_1_w, nullptr,+ nullptr, nullptr,+ model.mm_2_w, nullptr,+ FFN_GELU_ERF, -1);+ cb(projected, "after_projection", -1);+ }++ ggml_build_forward_expand(gf, projected);+ return gf;+}diff --git a/tools/mtmd/models/models.h b/tools/mtmd/models/models.hindex 2d7555da41..caed438ec5 100644--- a/tools/mtmd/models/models.h+++ b/tools/mtmd/models/models.h@@ -210,6 +210,11 @@ struct clip_graph_qwen3a : clip_graph {ggml_cgraph * build() override;};+struct clip_graph_mimo_audio : clip_graph {+ clip_graph_mimo_audio(clip_ctx * ctx, const clip_image_f32 & img) : clip_graph(ctx, img) {}+ ggml_cgraph * build() override;+};+struct clip_graph_kimik25 : clip_graph {clip_graph_kimik25(clip_ctx * ctx, const clip_image_f32 & img) : clip_graph(ctx, img) {}ggml_cgraph * build() override;diff --git a/tools/mtmd/mtmd-audio.cpp b/tools/mtmd/mtmd-audio.cppindex b72fd067a5..ed68951c01 100644--- a/tools/mtmd/mtmd-audio.cpp+++ b/tools/mtmd/mtmd-audio.cpp@@ -725,6 +725,72 @@ bool mtmd_audio_preprocessor_qwen3a::preprocess(const float * sareturn true;}+//+// mtmd_audio_preprocessor_mimo_audio+//+// Matches torchaudio.transforms.MelSpectrogram(power=1.0, center=True) followed by+// log(clip(spec, min=1e-7)): HTK mel scale, no Slaney area norm, magnitude (not power)+// spectrogram, natural log, reflect-padded by n_fft/2 on each side.+//++void mtmd_audio_preprocessor_mimo_audio::initialize() {+ cache.fill_sin_cos_table(hparams.audio_n_fft);+ cache.fill_hann_window(hparams.audio_window_len, true);+ cache.fill_mel_filterbank_matrix(+ hparams.n_mel_bins, hparams.audio_n_fft, hparams.audio_sample_rate,+ 0.0f, hparams.audio_sample_rate / 2.0f,+ /*slaney_area_norm=*/ false,+ /*scale=*/ 1.0f,+ /*use_htk=*/ true+ );+}++bool mtmd_audio_preprocessor_mimo_audio::preprocess(const float * samples,+ size_t n_samples,+ std::vector<mtmd_audio_mel> & output) {+ if (n_samples == 0) {+ return false;+ }++ GGML_ASSERT(!cache.sin_vals.empty());+ GGML_ASSERT(!cache.cos_vals.empty());+ GGML_ASSERT(!cache.filters.data.empty());++ const int pad = hparams.audio_n_fft / 2;++ std::vector<float> padded(n_samples + 2 * pad, 0.0f);+ for (int i = 0; i < pad; i++) {+ int src = pad - i;+ padded[i] = (src < (int)n_samples) ? samples[src] : 0.0f;+ }+ std::copy(samples, samples + n_samples, padded.begin() + pad);+ for (int i = 0; i < pad; i++) {+ int src = (int)n_samples - 2 - i;+ padded[n_samples + pad + i] = (src >= 0) ? samples[src] : 0.0f;+ }++ filter_params params;+ params.n_mel = hparams.n_mel_bins;+ params.n_fft_bins = 1 + (hparams.audio_n_fft / 2);+ params.hann_window_size = hparams.audio_window_len;+ params.hop_length = hparams.audio_hop_len;+ params.sample_rate = hparams.audio_sample_rate;+ params.no_padding = true; // reflect padding already applied above+ params.use_natural_log = true;+ params.use_magnitude = true;+ params.mel_floor = 1e-7f;+ params.norm_per_feature = false;++ mtmd_audio_mel out;+ bool ok = log_mel_spectrogram(padded.data(), (int)padded.size(), 4, params, cache, out);+ if (!ok) {+ return false;+ }++ output.push_back(std::move(out));+ return true;+}+//// mtmd_audio_preprocessor_conformer//diff --git a/tools/mtmd/mtmd-audio.h b/tools/mtmd/mtmd-audio.hindex ad96bd847c..d8ec72b9d5 100644--- a/tools/mtmd/mtmd-audio.h+++ b/tools/mtmd/mtmd-audio.h@@ -111,6 +111,15 @@ struct mtmd_audio_preprocessor_qwen3a : mtmd_audio_preprocessor {mtmd_audio_cache cache;};+struct mtmd_audio_preprocessor_mimo_audio : mtmd_audio_preprocessor {+ mtmd_audio_preprocessor_mimo_audio(const clip_ctx * ctx) : mtmd_audio_preprocessor(ctx) {}+ void initialize() override;+ bool preprocess(const float * samples, size_t n_samples, std::vector<mtmd_audio_mel> & output) override;++ private:+ mtmd_audio_cache cache;+};+//// streaming ISTFT - converts spectrogram frames back to audio one frame at a time//diff --git a/tools/mtmd/mtmd.cpp b/tools/mtmd/mtmd.cppindex bb49b211ef..6e61cf3e52 100644--- a/tools/mtmd/mtmd.cpp+++ b/tools/mtmd/mtmd.cpp@@ -730,6 +730,12 @@ struct mtmd_context {aud_end = "<audio|>";audio_preproc = std::make_unique<mtmd_audio_preprocessor_gemma4ua>(ctx_a);} break;+ case PROJECTOR_TYPE_MIMO_AUDIO:+ {+ aud_beg = "<|mimo_audio_start|>";+ aud_end = "<|mimo_audio_end|>";+ audio_preproc = std::make_unique<mtmd_audio_preprocessor_mimo_audio>(ctx_a);+ } break;default:throw std::runtime_error(string_format("%s: unexpected audio projector type %d\n", __func__, proj));}