Upload folder using huggingface_hub

.gitattributes

@@ -33,3 +33,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+tekken.json filter=lfs diff=lfs merge=lfs -text

consolidated.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ee8b29ec0ff94e63eef233331ea7bc10545aab6a622e43dcf80781d08b8f6f7d
+size 28392883664

convert_voxtral_hf_to_mistral.py

@@ -0,0 +1,221 @@
+# coding=utf-8
+# Copyright 2025 HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import argparse
+import gc
+import json
+import os
+import re
+
+from safetensors.torch import save_file
+from safetensors.torch import safe_open
+from huggingface_hub import snapshot_download
+
+from transformers import VoxtralConfig
+
+# fmt: off
+STATE_DICT_MAPPING = {
+    r"^language_model\.lm_head":                                         r"output",
+    r"^language_model\.model\.norm":                                     r"norm",
+    r"^language_model\.model\.embed_tokens":                             r"tok_embeddings",
+    r"^language_model\.model\.layers\.(\d+)\.input_layernorm":           r"layers.\1.attention_norm",
+    r"^language_model\.model\.layers\.(\d+)\.post_attention_layernorm":  r"layers.\1.ffn_norm",
+    r"^language_model\.model\.layers\.(\d+)\.self_attn\.(q|k|v|o)_proj": r"layers.\1.attention.w\2",
+    r"^language_model\.model\.layers\.(\d+)\.mlp\.gate_proj":            r"layers.\1.feed_forward.w1",
+    r"^language_model\.model\.layers\.(\d+)\.mlp\.down_proj":            r"layers.\1.feed_forward.w2",
+    r"^language_model\.model\.layers\.(\d+)\.mlp\.up_proj":              r"layers.\1.feed_forward.w3",
+    r"language_model.model.embed_tokens":                                r"tok_embeddings",
+    r"audio_tower.conv1":                                                r"mm_whisper_embeddings.whisper_encoder.conv_layers.0" ,
+    r"audio_tower.conv2":                                                r"mm_whisper_embeddings.whisper_encoder.conv_layers.1" ,
+    r"audio_tower.layer_norm":                                           r"mm_whisper_embeddings.whisper_encoder.transformer.norm" ,
+    r"audio_tower.layers.(\d+).self_attn.(q|k|v)_proj":                  r"mm_whisper_embeddings.whisper_encoder.transformer.layers.\1.attention.w\2" ,
+    r"audio_tower.layers.(\d+).self_attn.out_proj":                      r"mm_whisper_embeddings.whisper_encoder.transformer.layers.\1.attention.wo" ,
+    r"audio_tower.layers.(\d+).self_attn_layer_norm":                    r"mm_whisper_embeddings.whisper_encoder.transformer.layers.\1.attention_norm" ,
+    r"audio_tower.layers.(\d+).fc(\d+)":                                 r"mm_whisper_embeddings.whisper_encoder.transformer.layers.\1.feed_forward.w\2" ,
+    r"audio_tower.layers.(\d+).final_layer_norm":                        r"mm_whisper_embeddings.whisper_encoder.transformer.layers.\1.ffn_norm" ,
+    r"multi_modal_projector.linear_1":                                   r"mm_whisper_embeddings.audio_language_projection.0" ,
+    r"multi_modal_projector.linear_2":                                   r"mm_whisper_embeddings.audio_language_projection.2" ,
+}
+# fmt: on
+
+SKIP_KEYS = ["audio_tower.embed_positions.weight"]
+
+def add_quantization_config(config, hf_config: VoxtralConfig):
+    quantization_config = hf_config.quantization_config
+    mistral_ignore = [] # keys to ignore in the quantization config
+    for hf_key in quantization_config["ignore"]:
+        mistral_key = map_hf_key_to_mistral(hf_key)
+        mistral_ignore.append(mistral_key)
+    quantization_config["ignore"] = mistral_ignore
+    config["quantization"] = quantization_config
+
+    return config
+
+def map_hf_key_to_mistral(hf_key):
+    """Map a key from HF format to Mistral format"""
+    for pattern, replacement in STATE_DICT_MAPPING.items():
+        new_key, n_replace = re.subn(pattern, replacement, hf_key)
+        if n_replace > 0:
+            return new_key.replace("weight_scale", "qscale_weight")
+    
+    # If no mapping found, return the original key
+    return hf_key.replace("weight_scale", "qscale_weight")
+
+
+def permute_for_mistral_rope(tensor, n_heads, dim1, dim2):
+    """Reverse the ROPE permutation to get back to Mistral format."""
+    tensor = tensor.view(n_heads, 2, dim1 // n_heads // 2, dim2)
+    tensor = tensor.transpose(1, 2)
+    tensor = tensor.reshape(dim1, dim2)
+    return tensor
+
+
+def convert_state_dict(hf_state_dict, config):
+    """Convert HF Voxtral state dict to Mistral format"""
+    mistral_dict = {}
+    
+    num_attention_heads = config["n_heads"]
+    hidden_size = config["dim"]
+    head_dim = config["head_dim"]
+    num_key_value_heads = config["n_kv_heads"]
+    key_value_dim = head_dim * num_key_value_heads
+    query_dim = head_dim * num_attention_heads
+    
+    for hf_key, tensor in hf_state_dict.items():
+        if hf_key in SKIP_KEYS:
+            continue
+        
+        mistral_key = map_hf_key_to_mistral(hf_key)
+            
+        if "language_model" in hf_key:
+            if hf_key.endswith("q_proj.weight"):
+                tensor = permute_for_mistral_rope(tensor, num_attention_heads, query_dim, hidden_size)
+            elif hf_key.endswith("q_proj.weight_scale") and tensor.size(0) == num_attention_heads:
+                tensor = permute_for_mistral_rope(tensor, num_attention_heads, query_dim, 1)
+            elif hf_key.endswith("k_proj.weight"):
+                tensor = permute_for_mistral_rope(tensor, num_key_value_heads, key_value_dim, hidden_size)
+            elif hf_key.endswith("k_proj.weight_scale") and tensor.size(0) == num_key_value_heads:
+                tensor = permute_for_mistral_rope(tensor, num_key_value_heads, key_value_dim, 1)
+
+        mistral_dict[mistral_key] = tensor
+       
+    return mistral_dict
+
+
+def write_model(
+    input_path_or_repo,
+    output_dir,
+    unquantized_model_path=None,
+):
+    print("Converting HF Voxtral model to Mistral format.")
+    os.makedirs(output_dir, exist_ok=True)
+
+    # Load the HF Voxtral model
+    print(f"Loading HF Voxtral model from {input_path_or_repo}...")
+    hf_config = VoxtralConfig.from_pretrained(input_path_or_repo)
+
+    if os.path.exists(input_path_or_repo):
+        local_path = input_path_or_repo
+    else:
+        local_path = snapshot_download(input_path_or_repo)
+    
+    # Convert config
+    if unquantized_model_path is not None:
+        if os.path.exists(unquantized_model_path):
+            unquantized_model_path = unquantized_model_path
+        else:
+            unquantized_model_path = snapshot_download(unquantized_model_path)
+        config_path = os.path.join(unquantized_model_path, "params.json")
+        with open(config_path, "r") as f:
+            config = json.load(f)
+        config = add_quantization_config(config, hf_config)
+
+        with open(os.path.join(output_dir, "params.json"), "w") as f:
+            json.dump(config, f, indent=2)
+    else:
+        raise ValueError(f"Unquantized model config not found for {unquantized_model_path}")
+    
+    # Convert state dict
+    print("Converting state dict...")
+    tensor_files = sorted([f for f in os.listdir(os.path.join(local_path)) if f.endswith(".safetensors")])
+
+    hf_state_dict = {}
+
+    for file in tensor_files:
+        file_path = os.path.join(local_path, file)
+        with safe_open(file_path, framework="pt", device="cuda") as f:
+            for key in f.keys():
+                hf_state_dict[key] = f.get_tensor(key)    
+
+    mistral_state_dict = convert_state_dict(hf_state_dict, config)
+   
+    # save the state dict
+    save_file(mistral_state_dict, os.path.join(output_dir, "consolidated.safetensors"))
+
+    del hf_state_dict, mistral_state_dict
+    gc.collect()
+    print("Model converted successfully.")
+
+def write_tokenizer(input_path_or_repo: str, output_dir: str):
+    """Extract and save the tokenizer from Voxtral model"""
+    from transformers import MistralCommonTokenizer
+    
+    print("Extracting tokenizer...")
+    tokenizer = MistralCommonTokenizer.from_pretrained(input_path_or_repo)
+    tokenizer.save_pretrained(output_dir)
+    print("Tokenizer saved successfully.")
+
+
+def main():
+    parser = argparse.ArgumentParser(description="Convert HF Voxtral weights to Mistral format")
+    parser.add_argument(
+        "--input_path_or_repo",
+        type=str,
+        default="RedHatAI/Voxtral-Small-24B-2507-FP8-dynamic",
+        help="Path or repo containing HF Voxtral model",
+    )
+    parser.add_argument(
+        "--output_dir",
+        type=str,
+        default="Voxtral-Small-24B-2507-FP8-dynamic-converted",
+        help="Location to write Mistral model and tokenizer",
+    )
+    parser.add_argument(
+        "--skip_tokenizer",
+        action="store_true",
+        help="Skip tokenizer conversion"
+    )
+    parser.add_argument(
+        "--unquantized_model_path",
+        type=str,
+        default="mistralai/Voxtral-Small-24B-2507",
+        help="Path to the unquantized model",
+    )
+    args = parser.parse_args()
+
+    write_model(
+        args.input_path_or_repo,
+        args.output_dir,
+        unquantized_model_path=args.unquantized_model_path,
+    )
+
+    if not args.skip_tokenizer:
+        write_tokenizer(
+            args.input_path_or_repo,
+            args.output_dir,
+        )
+
+
+if __name__ == "__main__":
+    main()

params.json

@@ -0,0 +1,434 @@
+{
+  "dim": 5120,
+  "n_layers": 40,
+  "head_dim": 128,
+  "hidden_dim": 32768,
+  "n_heads": 32,
+  "n_kv_heads": 8,
+  "rope_theta": 100000000.0,
+  "norm_eps": 1e-05,
+  "vocab_size": 131072,
+  "max_position_embeddings": 32768,
+  "multimodal": {
+    "whisper_model_args": {
+      "encoder_args": {
+        "dim": 1280,
+        "n_layers": 32,
+        "head_dim": 64,
+        "hidden_dim": 5120,
+        "n_heads": 20,
+        "vocab_size": 51866,
+        "max_source_positions": 1500,
+        "audio_encoding_args": {
+          "sampling_rate": 16000,
+          "num_mel_bins": 128,
+          "hop_length": 160,
+          "window_size": 400
+        }
+      },
+      "downsample_args": {
+        "downsample_factor": 4
+      }
+    }
+  },
+  "quantization": {
+    "config_groups": {
+      "group_0": {
+        "format": "float-quantized",
+        "input_activations": {
+          "actorder": null,
+          "block_structure": null,
+          "dynamic": true,
+          "group_size": null,
+          "num_bits": 8,
+          "observer": null,
+          "observer_kwargs": {},
+          "strategy": "token",
+          "symmetric": true,
+          "type": "float"
+        },
+        "output_activations": null,
+        "targets": [
+          "Linear"
+        ],
+        "weights": {
+          "actorder": null,
+          "block_structure": null,
+          "dynamic": false,
+          "group_size": null,
+          "num_bits": 8,
+          "observer": "mse",
+          "observer_kwargs": {},
+          "strategy": "channel",
+          "symmetric": true,
+          "type": "float"
+        }
+      }
+    },
+    "format": "float-quantized",
+    "global_compression_ratio": null,
+    "ignore": [
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.0.attention.wk",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.0.attention.wv",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.0.attention.wq",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.0.attention.wo",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.0.feed_forward.w1",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.0.feed_forward.w2",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.1.attention.wk",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.1.attention.wv",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.1.attention.wq",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.1.attention.wo",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.1.feed_forward.w1",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.1.feed_forward.w2",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.2.attention.wk",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.2.attention.wv",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.2.attention.wq",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.2.attention.wo",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.2.feed_forward.w1",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.2.feed_forward.w2",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.3.attention.wk",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.3.attention.wv",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.3.attention.wq",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.3.attention.wo",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.3.feed_forward.w1",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.3.feed_forward.w2",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.4.attention.wk",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.4.attention.wv",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.4.attention.wq",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.4.attention.wo",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.4.feed_forward.w1",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.4.feed_forward.w2",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.5.attention.wk",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.5.attention.wv",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.5.attention.wq",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.5.attention.wo",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.5.feed_forward.w1",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.5.feed_forward.w2",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.6.attention.wk",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.6.attention.wv",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.6.attention.wq",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.6.attention.wo",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.6.feed_forward.w1",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.6.feed_forward.w2",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.7.attention.wk",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.7.attention.wv",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.7.attention.wq",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.7.attention.wo",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.7.feed_forward.w1",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.7.feed_forward.w2",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.8.attention.wk",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.8.attention.wv",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.8.attention.wq",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.8.attention.wo",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.8.feed_forward.w1",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.8.feed_forward.w2",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.9.attention.wk",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.9.attention.wv",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.9.attention.wq",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.9.attention.wo",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.9.feed_forward.w1",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.9.feed_forward.w2",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.10.attention.wk",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.10.attention.wv",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.10.attention.wq",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.10.attention.wo",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.10.feed_forward.w1",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.10.feed_forward.w2",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.11.attention.wk",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.11.attention.wv",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.11.attention.wq",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.11.attention.wo",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.11.feed_forward.w1",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.11.feed_forward.w2",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.12.attention.wk",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.12.attention.wv",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.12.attention.wq",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.12.attention.wo",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.12.feed_forward.w1",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.12.feed_forward.w2",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.13.attention.wk",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.13.attention.wv",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.13.attention.wq",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.13.attention.wo",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.13.feed_forward.w1",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.13.feed_forward.w2",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.14.attention.wk",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.14.attention.wv",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.14.attention.wq",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.14.attention.wo",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.14.feed_forward.w1",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.14.feed_forward.w2",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.15.attention.wk",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.15.attention.wv",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.15.attention.wq",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.15.attention.wo",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.15.feed_forward.w1",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.15.feed_forward.w2",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.16.attention.wk",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.16.attention.wv",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.16.attention.wq",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.16.attention.wo",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.16.feed_forward.w1",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.16.feed_forward.w2",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.17.attention.wk",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.17.attention.wv",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.17.attention.wq",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.17.attention.wo",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.17.feed_forward.w1",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.17.feed_forward.w2",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.18.attention.wk",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.18.attention.wv",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.18.attention.wq",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.18.attention.wo",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.18.feed_forward.w1",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.18.feed_forward.w2",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.19.attention.wk",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.19.attention.wv",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.19.attention.wq",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.19.attention.wo",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.19.feed_forward.w1",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.19.feed_forward.w2",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.20.attention.wk",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.20.attention.wv",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.20.attention.wq",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.20.attention.wo",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.20.feed_forward.w1",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.20.feed_forward.w2",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.21.attention.wk",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.21.attention.wv",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.21.attention.wq",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.21.attention.wo",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.21.feed_forward.w1",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.21.feed_forward.w2",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.22.attention.wk",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.22.attention.wv",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.22.attention.wq",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.22.attention.wo",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.22.feed_forward.w1",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.22.feed_forward.w2",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.23.attention.wk",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.23.attention.wv",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.23.attention.wq",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.23.attention.wo",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.23.feed_forward.w1",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.23.feed_forward.w2",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.24.attention.wk",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.24.attention.wv",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.24.attention.wq",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.24.attention.wo",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.24.feed_forward.w1",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.24.feed_forward.w2",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.25.attention.wk",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.25.attention.wv",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.25.attention.wq",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.25.attention.wo",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.25.feed_forward.w1",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.25.feed_forward.w2",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.26.attention.wk",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.26.attention.wv",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.26.attention.wq",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.26.attention.wo",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.26.feed_forward.w1",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.26.feed_forward.w2",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.27.attention.wk",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.27.attention.wv",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.27.attention.wq",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.27.attention.wo",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.27.feed_forward.w1",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.27.feed_forward.w2",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.28.attention.wk",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.28.attention.wv",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.28.attention.wq",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.28.attention.wo",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.28.feed_forward.w1",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.28.feed_forward.w2",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.29.attention.wk",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.29.attention.wv",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.29.attention.wq",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.29.attention.wo",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.29.feed_forward.w1",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.29.feed_forward.w2",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.30.attention.wk",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.30.attention.wv",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.30.attention.wq",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.30.attention.wo",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.30.feed_forward.w1",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.30.feed_forward.w2",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.31.attention.wk",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.31.attention.wv",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.31.attention.wq",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.31.attention.wo",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.31.feed_forward.w1",
+      "mm_whisper_embeddings.whisper_encoder.transformer.layers.31.feed_forward.w2",
+      "layers.0.attention.wq",
+      "layers.0.attention.wk",
+      "layers.0.attention.wv",
+      "layers.0.attention.wo",
+      "layers.1.attention.wq",
+      "layers.1.attention.wk",
+      "layers.1.attention.wv",
+      "layers.1.attention.wo",
+      "layers.2.attention.wq",
+      "layers.2.attention.wk",
+      "layers.2.attention.wv",
+      "layers.2.attention.wo",
+      "layers.3.attention.wq",
+      "layers.3.attention.wk",
+      "layers.3.attention.wv",
+      "layers.3.attention.wo",
+      "layers.4.attention.wq",
+      "layers.4.attention.wk",
+      "layers.4.attention.wv",
+      "layers.4.attention.wo",
+      "layers.5.attention.wq",
+      "layers.5.attention.wk",
+      "layers.5.attention.wv",
+      "layers.5.attention.wo",
+      "layers.6.attention.wq",
+      "layers.6.attention.wk",
+      "layers.6.attention.wv",
+      "layers.6.attention.wo",
+      "layers.7.attention.wq",
+      "layers.7.attention.wk",
+      "layers.7.attention.wv",
+      "layers.7.attention.wo",
+      "layers.8.attention.wq",
+      "layers.8.attention.wk",
+      "layers.8.attention.wv",
+      "layers.8.attention.wo",
+      "layers.9.attention.wq",
+      "layers.9.attention.wk",
+      "layers.9.attention.wv",
+      "layers.9.attention.wo",
+      "layers.10.attention.wq",
+      "layers.10.attention.wk",
+      "layers.10.attention.wv",
+      "layers.10.attention.wo",
+      "layers.11.attention.wq",
+      "layers.11.attention.wk",
+      "layers.11.attention.wv",
+      "layers.11.attention.wo",
+      "layers.12.attention.wq",
+      "layers.12.attention.wk",
+      "layers.12.attention.wv",
+      "layers.12.attention.wo",
+      "layers.13.attention.wq",
+      "layers.13.attention.wk",
+      "layers.13.attention.wv",
+      "layers.13.attention.wo",
+      "layers.14.attention.wq",
+      "layers.14.attention.wk",
+      "layers.14.attention.wv",
+      "layers.14.attention.wo",
+      "layers.15.attention.wq",
+      "layers.15.attention.wk",
+      "layers.15.attention.wv",
+      "layers.15.attention.wo",
+      "layers.16.attention.wq",
+      "layers.16.attention.wk",
+      "layers.16.attention.wv",
+      "layers.16.attention.wo",
+      "layers.17.attention.wq",
+      "layers.17.attention.wk",
+      "layers.17.attention.wv",
+      "layers.17.attention.wo",
+      "layers.18.attention.wq",
+      "layers.18.attention.wk",
+      "layers.18.attention.wv",
+      "layers.18.attention.wo",
+      "layers.19.attention.wq",
+      "layers.19.attention.wk",
+      "layers.19.attention.wv",
+      "layers.19.attention.wo",
+      "layers.20.attention.wq",
+      "layers.20.attention.wk",
+      "layers.20.attention.wv",
+      "layers.20.attention.wo",
+      "layers.21.attention.wq",
+      "layers.21.attention.wk",
+      "layers.21.attention.wv",
+      "layers.21.attention.wo",
+      "layers.22.attention.wq",
+      "layers.22.attention.wk",
+      "layers.22.attention.wv",
+      "layers.22.attention.wo",
+      "layers.23.attention.wq",
+      "layers.23.attention.wk",
+      "layers.23.attention.wv",
+      "layers.23.attention.wo",
+      "layers.24.attention.wq",
+      "layers.24.attention.wk",
+      "layers.24.attention.wv",
+      "layers.24.attention.wo",
+      "layers.25.attention.wq",
+      "layers.25.attention.wk",
+      "layers.25.attention.wv",
+      "layers.25.attention.wo",
+      "layers.26.attention.wq",
+      "layers.26.attention.wk",
+      "layers.26.attention.wv",
+      "layers.26.attention.wo",
+      "layers.27.attention.wq",
+      "layers.27.attention.wk",
+      "layers.27.attention.wv",
+      "layers.27.attention.wo",
+      "layers.28.attention.wq",
+      "layers.28.attention.wk",
+      "layers.28.attention.wv",
+      "layers.28.attention.wo",
+      "layers.29.attention.wq",
+      "layers.29.attention.wk",
+      "layers.29.attention.wv",
+      "layers.29.attention.wo",
+      "layers.30.attention.wq",
+      "layers.30.attention.wk",
+      "layers.30.attention.wv",
+      "layers.30.attention.wo",
+      "layers.31.attention.wq",
+      "layers.31.attention.wk",
+      "layers.31.attention.wv",
+      "layers.31.attention.wo",
+      "layers.32.attention.wq",
+      "layers.32.attention.wk",
+      "layers.32.attention.wv",
+      "layers.32.attention.wo",
+      "layers.33.attention.wq",
+      "layers.33.attention.wk",
+      "layers.33.attention.wv",
+      "layers.33.attention.wo",
+      "layers.34.attention.wq",
+      "layers.34.attention.wk",
+      "layers.34.attention.wv",
+      "layers.34.attention.wo",
+      "layers.35.attention.wq",
+      "layers.35.attention.wk",
+      "layers.35.attention.wv",
+      "layers.35.attention.wo",
+      "layers.36.attention.wq",
+      "layers.36.attention.wk",
+      "layers.36.attention.wv",
+      "layers.36.attention.wo",
+      "layers.37.attention.wq",
+      "layers.37.attention.wk",
+      "layers.37.attention.wv",
+      "layers.37.attention.wo",
+      "layers.38.attention.wq",
+      "layers.38.attention.wk",
+      "layers.38.attention.wv",
+      "layers.38.attention.wo",
+      "layers.39.attention.wq",
+      "layers.39.attention.wk",
+      "layers.39.attention.wv",
+      "layers.39.attention.wo",
+      "output",
+      "mm_whisper_embeddings.audio_language_projection.0",
+      "mm_whisper_embeddings.audio_language_projection.2"
+    ],
+    "kv_cache_scheme": null,
+    "quant_method": "compressed-tensors",
+    "quantization_status": "compressed",
+    "sparsity_config": {},
+    "transform_config": {},
+    "version": "0.11.1.a20250923"
+  }
+}

tekken.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4aaf3836c2a5332f029ce85a7a62255c966f47b6797ef81dedd0ade9c862e4a8
+size 14894206