Upload 8 files

.gitattributes

@@ -7,3 +7,6 @@ models/experiments/ckpt/nar-len-llama-9/fp32.pth filter=lfs diff=lfs merge=lfs -
 models/ckpt/ar+nar-llama-8/fp32.sft filter=lfs diff=lfs merge=lfs -text
 models/ckpt/ar+nar-retnet-8/fp32.sft filter=lfs diff=lfs merge=lfs -text
 models/ckpt/ar+nar-tts+stt-llama-8/fp32.sft filter=lfs diff=lfs merge=lfs -text
+loras/ckpt/lora-glados-r128-a128/lora.sft filter=lfs diff=lfs merge=lfs -text
+loras/ckpt/lora-max-r128-a128/lora.sft filter=lfs diff=lfs merge=lfs -text
+loras/ckpt/lora-sam-r128-a128/lora.sft filter=lfs diff=lfs merge=lfs -text

loras/ckpt/lora-glados-r128-a128/lora.sft

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6e9bec9574b8a074c9a87f8997b27706f4f3611efa12d26dc9763cd2bca9eb9e
+size 66076312

loras/ckpt/lora-max-r128-a128/lora.sft

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0e8065b17f981dba303af2e5c92275902a45e8e38e9eea578f0a9e4b50690f1e
+size 66076312

loras/ckpt/lora-sam-r128-a128/lora.sft

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:433ada1d6f9cba5ebce7917bb51df09f050f4d221afdc4cdad72964b6624f9fc
+size 66076312

loras/config.lora.glados.yaml

@@ -0,0 +1,130 @@
+sample_rate: 24_000
+audio_backend: "vocos"
+
+models:
+- name: "ar+nar-tts+stt"
+  size: "full"
+  resp_levels: 8
+  prom_levels: 8
+  tasks: 9
+  langs: 2
+  tones: 1
+  arch_type: llama
+  training: False
+  version: 5
+  attention: auto
+  dropout: 0.1
+  #loss_factors:
+  #  text: 0.01
+  #  prom: 0.5
+  #  resp: 1.0
+  capabilities: ["ar", "nar"]
+  experimental:
+    p_rvq_levels: "auto"
+    audio_embedding_sums: True
+    unified_position_ids: False
+    split_classifiers: True
+    # 
+    causal_size: 1
+    interleave: False
+    rvq_level_range: []
+    tie_classifier_to_embedding: False
+
+loras:
+- name : "lora-glados"
+  rank: 128
+  alpha: 128
+  training: True
+  rvq_levels: []
+
+hyperparameters:
+  batch_size: 32
+  gradient_accumulation_steps: 8
+  gradient_clipping: 1.0
+  warmup_steps: 10
+
+  optimizer: Prodigy
+  learning_rate: 1.0
+  torch_optimizer: True
+  
+  scheduler: "" # ScheduleFree
+  torch_scheduler: True
+
+evaluation:
+  batch_size: 4
+  frequency: 250
+  size: 4
+  
+  steps: 500
+  ar_temperature: 1.0
+  nar_temperature: 0.0
+
+trainer:
+  iterations: 1_000_000  
+  save_frequency: 250
+  keep_last_checkpoints: 4
+
+  resize_modules: True
+  
+  check_for_oom: False
+  gradient_checkpointing: True
+
+  weight_dtype: bfloat16
+  amp: True
+
+  backend: deepspeed
+  deepspeed:
+    inferencing: False
+    amp: False
+
+  load_webui: False
+
+inference:
+  backend: local
+  normalize: False
+
+  weight_dtype: bfloat16
+  amp: True
+
+optimizations:
+  injects: False
+  replace: True
+
+  linear: False
+  embedding: False
+  optimizers: True
+
+  bitsandbytes: False
+  dadaptation: False
+  bitnet: False
+  fp8: False
+
+dataset:
+  use_hdf5: True
+  hdf5_flag: r
+  
+  use_metadata: True
+  validate: True
+
+  workers: 1
+  cache: True
+
+  duration_range: [3.0, 12.0]
+
+  random_utterance: 1.0
+  max_prompts: 1
+  prompt_duration_range: [3.0, 3.0]
+  
+  max_resps: 1
+  p_resp_append: 0.25
+
+  sample_type: path # path # speaker
+  sample_order: duration
+  sample_max_duration_batch: 300
+  sample_shuffle: False
+
+  tasks_list: [ "tts", "stt" ]
+
+  training: []
+  validation: []
+  noise: []

loras/config.lora.max.yaml

@@ -0,0 +1,130 @@
+sample_rate: 24_000
+audio_backend: "vocos"
+
+models:
+- name: "ar+nar-tts+stt"
+  size: "full"
+  resp_levels: 8
+  prom_levels: 8
+  tasks: 9
+  langs: 2
+  tones: 1
+  arch_type: llama
+  training: False
+  version: 5
+  attention: auto
+  dropout: 0.1
+  #loss_factors:
+  #  text: 0.01
+  #  prom: 0.5
+  #  resp: 1.0
+  capabilities: ["ar", "nar"]
+  experimental:
+    p_rvq_levels: "auto"
+    audio_embedding_sums: True
+    unified_position_ids: False
+    split_classifiers: True
+    # 
+    causal_size: 1
+    interleave: False
+    rvq_level_range: []
+    tie_classifier_to_embedding: False
+
+loras:
+- name : "lora-max"
+  rank: 128
+  alpha: 128
+  training: True
+  rvq_levels: []
+
+hyperparameters:
+  batch_size: 32
+  gradient_accumulation_steps: 8
+  gradient_clipping: 1.0
+  warmup_steps: 10
+
+  optimizer: Prodigy
+  learning_rate: 1.0
+  torch_optimizer: True
+  
+  scheduler: "" # ScheduleFree
+  torch_scheduler: True
+
+evaluation:
+  batch_size: 4
+  frequency: 250
+  size: 4
+  
+  steps: 500
+  ar_temperature: 1.0
+  nar_temperature: 0.0
+
+trainer:
+  iterations: 1_000_000  
+  save_frequency: 250
+  keep_last_checkpoints: 4
+
+  resize_modules: True
+  
+  check_for_oom: False
+  gradient_checkpointing: True
+
+  weight_dtype: bfloat16
+  amp: True
+
+  backend: deepspeed
+  deepspeed:
+    inferencing: False
+    amp: False
+
+  load_webui: False
+
+inference:
+  backend: local
+  normalize: False
+
+  weight_dtype: bfloat16
+  amp: True
+
+optimizations:
+  injects: False
+  replace: True
+
+  linear: False
+  embedding: False
+  optimizers: True
+
+  bitsandbytes: False
+  dadaptation: False
+  bitnet: False
+  fp8: False
+
+dataset:
+  use_hdf5: True
+  hdf5_flag: r
+  
+  use_metadata: True
+  validate: True
+
+  workers: 1
+  cache: True
+
+  duration_range: [3.0, 12.0]
+
+  random_utterance: 1.0
+  max_prompts: 1
+  prompt_duration_range: [3.0, 3.0]
+  
+  max_resps: 1
+  p_resp_append: 0.25
+
+  sample_type: path # path # speaker
+  sample_order: duration
+  sample_max_duration_batch: 300
+  sample_shuffle: False
+
+  tasks_list: [ "tts", "stt" ]
+
+  training: []
+  validation: []
+  noise: []

loras/config.lora.sam.yaml

@@ -0,0 +1,130 @@
+sample_rate: 24_000
+audio_backend: "vocos"
+
+models:
+- name: "ar+nar-tts+stt"
+  size: "full"
+  resp_levels: 8
+  prom_levels: 8
+  tasks: 9
+  langs: 2
+  tones: 1
+  arch_type: llama
+  training: False
+  version: 5
+  attention: auto
+  dropout: 0.1
+  #loss_factors:
+  #  text: 0.01
+  #  prom: 0.5
+  #  resp: 1.0
+  capabilities: ["ar", "nar"]
+  experimental:
+    p_rvq_levels: "auto"
+    audio_embedding_sums: True
+    unified_position_ids: False
+    split_classifiers: True
+    # 
+    causal_size: 1
+    interleave: False
+    rvq_level_range: []
+    tie_classifier_to_embedding: False
+
+loras:
+- name : "lora-sam"
+  rank: 128
+  alpha: 128
+  training: True
+  rvq_levels: []
+
+hyperparameters:
+  batch_size: 32
+  gradient_accumulation_steps: 8
+  gradient_clipping: 1.0
+  warmup_steps: 10
+
+  optimizer: Prodigy
+  learning_rate: 1.0
+  torch_optimizer: True
+  
+  scheduler: "" # ScheduleFree
+  torch_scheduler: True
+
+evaluation:
+  batch_size: 4
+  frequency: 250
+  size: 4
+  
+  steps: 500
+  ar_temperature: 1.0
+  nar_temperature: 0.0
+
+trainer:
+  iterations: 1_000_000  
+  save_frequency: 250
+  keep_last_checkpoints: 4
+
+  resize_modules: True
+  
+  check_for_oom: False
+  gradient_checkpointing: True
+
+  weight_dtype: bfloat16
+  amp: True
+
+  backend: deepspeed
+  deepspeed:
+    inferencing: False
+    amp: False
+
+  load_webui: False
+
+inference:
+  backend: local
+  normalize: False
+
+  weight_dtype: bfloat16
+  amp: True
+
+optimizations:
+  injects: False
+  replace: True
+
+  linear: False
+  embedding: False
+  optimizers: True
+
+  bitsandbytes: False
+  dadaptation: False
+  bitnet: False
+  fp8: False
+
+dataset:
+  use_hdf5: True
+  hdf5_flag: r
+  
+  use_metadata: True
+  validate: True
+
+  workers: 1
+  cache: True
+
+  duration_range: [3.0, 12.0]
+
+  random_utterance: 1.0
+  max_prompts: 1
+  prompt_duration_range: [3.0, 3.0]
+  
+  max_resps: 1
+  p_resp_append: 0.25
+
+  sample_type: path # path # speaker
+  sample_order: duration
+  sample_max_duration_batch: 300
+  sample_shuffle: False
+
+  tasks_list: [ "tts", "stt" ]
+
+  training: []
+  validation: []
+  noise: []

models/ckpt/ar+nar-tts+stt-llama-8/fp32.sft

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:cbf25cb391bbbb79ee05b5ecb9b219dce8785d9561473001e0e041a70d0e634b
+oid sha256:fc995fd0c81f7540518cd41a2c0482eb81f03a86c5d8f58dfd4aca4ec5fead3d
 size 456272602

models/config.llama.tts+stt.yaml

@@ -20,14 +20,15 @@ models:
   #  resp: 1.0
   capabilities: ["ar", "nar"]
   experimental:
-    audio_embedding_sums: False
-    interleave: False
+    p_rvq_levels: "auto"
+    audio_embedding_sums: True
     unified_position_ids: False
-    rvq_level_range: []
     split_classifiers: True
-    tie_classifier_to_embedding: False
+    # 
     causal_size: 1
-    p_rvq_levels: "auto"
+    interleave: False
+    rvq_level_range: []
+    tie_classifier_to_embedding: False
 
 #loras:
 #- name : "lora"
@@ -99,13 +100,6 @@ optimizations:
   fp8: False
 
 dataset:
-  speaker_name_getter: "lambda p: f'{p.parts[-3]}_{p.parts[-2]}'"
-  speaker_group_getter: "lambda p: f'{p.parts[-3]}'"
-  speaker_languages:
-    ja: [
-      "housamo"
-    ]
-
   use_hdf5: True
   hdf5_flag: r
   
@@ -129,6 +123,8 @@ dataset:
   sample_max_duration_batch: 300
   sample_shuffle: False
 
+  tasks_list: [ "tts", "stt" ]
+
   training: []
   validation: []
   noise: []