Add Gemma 4 model sampling defaults (#4838)

Add per-model YAML configs and MODEL_NAME_MAPPING entries for all 8
Gemma 4 models (4 instruct + 4 base):
- gemma-4-31B-it / gemma-4-31B
- gemma-4-26B-A4B-it / gemma-4-26B-A4B
- gemma-4-E2B-it / gemma-4-E2B
- gemma-4-E4B-it / gemma-4-E4B

GGUF variants (only for -it models) resolve via the gemma-4 family
entry in inference_defaults.json.

Sampling defaults: temperature=1.0, top_p=0.95, top_k=64, min_p=0.0,
no repetition or presence penalty. Matches gemma-3n and gemma-3.

studio/backend/assets/configs/inference_defaults.json

@@ -93,6 +93,14 @@
       "min_p": 0.0,
       "repetition_penalty": 1.0
     },
+    "gemma-4": {
+      "temperature": 1.0,
+      "top_p": 0.95,
+      "top_k": 64,
+      "min_p": 0.0,
+      "repetition_penalty": 1.0,
+      "presence_penalty": 0.0
+    },
     "gemma-3n": {
       "temperature": 1.0,
       "top_p": 0.95,
@@ -366,7 +374,7 @@
     "qwen2.5-coder", "qwen2.5-vl", "qwen2.5-omni", "qwen2.5-math", "qwen2.5",
     "qwen2-vl", "qwen2",
     "qwq",
-    "gemma-3n", "gemma-3", "medgemma", "gemma-2",
+    "gemma-4", "gemma-3n", "gemma-3", "medgemma", "gemma-2",
     "llama-4", "llama-3.3", "llama-3.2", "llama-3.1", "llama-3",
     "phi-4", "phi-3",
     "mistral-nemo", "mistral-small", "mistral-large", "magistral", "ministral",

studio/backend/assets/configs/model_defaults/gemma/unsloth_gemma-4-26B-A4B-it.yaml

@@ -0,0 +1,47 @@
+# Model defaults for unsloth/gemma-4-26B-A4B-it
+# Also applies to: google/gemma-4-26B-A4B-it, unsloth/gemma-4-26B-A4B-it-GGUF
+
+training:
+  trust_remote_code: false
+  max_seq_length: 2048
+  num_epochs: 0
+  learning_rate: 2e-4
+  batch_size: 2
+  gradient_accumulation_steps: 4
+  warmup_steps: 5
+  max_steps: 30
+  save_steps: 30
+  weight_decay: 0.001
+  random_seed: 3407
+  packing: false
+  train_on_completions: true
+  gradient_checkpointing: "unsloth"
+  optim: "adamw_8bit"
+  lr_scheduler_type: "linear"
+
+lora:
+  lora_r: 8
+  lora_alpha: 8
+  lora_dropout: 0.0
+  target_modules:
+    - "all-linear"
+  use_rslora: false
+  use_loftq: false
+  finetune_vision_layers: true
+  finetune_language_layers: true
+  finetune_attention_modules: true
+  finetune_mlp_modules: true
+
+logging:
+  enable_wandb: false
+  wandb_project: "llm-finetuning"
+  enable_tensorboard: false
+  tensorboard_dir: "runs"
+  log_frequency: 10
+
+inference:
+  trust_remote_code: false
+  temperature: 1.0
+  top_p: 0.95
+  top_k: 64
+  min_p: 0.0

studio/backend/assets/configs/model_defaults/gemma/unsloth_gemma-4-26B-A4B.yaml

@@ -0,0 +1,47 @@
+# Model defaults for unsloth/gemma-4-26B-A4B (base/pretrained)
+# Also applies to: google/gemma-4-26B-A4B
+
+training:
+  trust_remote_code: false
+  max_seq_length: 2048
+  num_epochs: 0
+  learning_rate: 2e-4
+  batch_size: 2
+  gradient_accumulation_steps: 4
+  warmup_steps: 5
+  max_steps: 30
+  save_steps: 30
+  weight_decay: 0.001
+  random_seed: 3407
+  packing: false
+  train_on_completions: true
+  gradient_checkpointing: "unsloth"
+  optim: "adamw_8bit"
+  lr_scheduler_type: "linear"
+
+lora:
+  lora_r: 8
+  lora_alpha: 8
+  lora_dropout: 0.0
+  target_modules:
+    - "all-linear"
+  use_rslora: false
+  use_loftq: false
+  finetune_vision_layers: true
+  finetune_language_layers: true
+  finetune_attention_modules: true
+  finetune_mlp_modules: true
+
+logging:
+  enable_wandb: false
+  wandb_project: "llm-finetuning"
+  enable_tensorboard: false
+  tensorboard_dir: "runs"
+  log_frequency: 10
+
+inference:
+  trust_remote_code: false
+  temperature: 1.0
+  top_p: 0.95
+  top_k: 64
+  min_p: 0.0

studio/backend/assets/configs/model_defaults/gemma/unsloth_gemma-4-31B-it.yaml

@@ -0,0 +1,47 @@
+# Model defaults for unsloth/gemma-4-31B-it
+# Also applies to: google/gemma-4-31B-it, unsloth/gemma-4-31B-it-GGUF
+
+training:
+  trust_remote_code: false
+  max_seq_length: 2048
+  num_epochs: 0
+  learning_rate: 2e-4
+  batch_size: 2
+  gradient_accumulation_steps: 4
+  warmup_steps: 5
+  max_steps: 30
+  save_steps: 30
+  weight_decay: 0.001
+  random_seed: 3407
+  packing: false
+  train_on_completions: true
+  gradient_checkpointing: "unsloth"
+  optim: "adamw_8bit"
+  lr_scheduler_type: "linear"
+
+lora:
+  lora_r: 8
+  lora_alpha: 8
+  lora_dropout: 0.0
+  target_modules:
+    - "all-linear"
+  use_rslora: false
+  use_loftq: false
+  finetune_vision_layers: true
+  finetune_language_layers: true
+  finetune_attention_modules: true
+  finetune_mlp_modules: true
+
+logging:
+  enable_wandb: false
+  wandb_project: "llm-finetuning"
+  enable_tensorboard: false
+  tensorboard_dir: "runs"
+  log_frequency: 10
+
+inference:
+  trust_remote_code: false
+  temperature: 1.0
+  top_p: 0.95
+  top_k: 64
+  min_p: 0.0

studio/backend/assets/configs/model_defaults/gemma/unsloth_gemma-4-31B.yaml

@@ -0,0 +1,47 @@
+# Model defaults for unsloth/gemma-4-31B (base/pretrained)
+# Also applies to: google/gemma-4-31B
+
+training:
+  trust_remote_code: false
+  max_seq_length: 2048
+  num_epochs: 0
+  learning_rate: 2e-4
+  batch_size: 2
+  gradient_accumulation_steps: 4
+  warmup_steps: 5
+  max_steps: 30
+  save_steps: 30
+  weight_decay: 0.001
+  random_seed: 3407
+  packing: false
+  train_on_completions: true
+  gradient_checkpointing: "unsloth"
+  optim: "adamw_8bit"
+  lr_scheduler_type: "linear"
+
+lora:
+  lora_r: 8
+  lora_alpha: 8
+  lora_dropout: 0.0
+  target_modules:
+    - "all-linear"
+  use_rslora: false
+  use_loftq: false
+  finetune_vision_layers: true
+  finetune_language_layers: true
+  finetune_attention_modules: true
+  finetune_mlp_modules: true
+
+logging:
+  enable_wandb: false
+  wandb_project: "llm-finetuning"
+  enable_tensorboard: false
+  tensorboard_dir: "runs"
+  log_frequency: 10
+
+inference:
+  trust_remote_code: false
+  temperature: 1.0
+  top_p: 0.95
+  top_k: 64
+  min_p: 0.0

studio/backend/assets/configs/model_defaults/gemma/unsloth_gemma-4-E2B-it.yaml

@@ -0,0 +1,47 @@
+# Model defaults for unsloth/gemma-4-E2B-it
+# Also applies to: google/gemma-4-E2B-it, unsloth/gemma-4-E2B-it-GGUF
+
+training:
+  trust_remote_code: false
+  max_seq_length: 2048
+  num_epochs: 0
+  learning_rate: 2e-4
+  batch_size: 2
+  gradient_accumulation_steps: 4
+  warmup_steps: 5
+  max_steps: 30
+  save_steps: 30
+  weight_decay: 0.001
+  random_seed: 3407
+  packing: false
+  train_on_completions: true
+  gradient_checkpointing: "unsloth"
+  optim: "adamw_8bit"
+  lr_scheduler_type: "linear"
+
+lora:
+  lora_r: 8
+  lora_alpha: 8
+  lora_dropout: 0.0
+  target_modules:
+    - "all-linear"
+  use_rslora: false
+  use_loftq: false
+  finetune_vision_layers: true
+  finetune_language_layers: true
+  finetune_attention_modules: true
+  finetune_mlp_modules: true
+
+logging:
+  enable_wandb: false
+  wandb_project: "llm-finetuning"
+  enable_tensorboard: false
+  tensorboard_dir: "runs"
+  log_frequency: 10
+
+inference:
+  trust_remote_code: false
+  temperature: 1.0
+  top_p: 0.95
+  top_k: 64
+  min_p: 0.0

studio/backend/assets/configs/model_defaults/gemma/unsloth_gemma-4-E2B.yaml

@@ -0,0 +1,47 @@
+# Model defaults for unsloth/gemma-4-E2B (base/pretrained)
+# Also applies to: google/gemma-4-E2B
+
+training:
+  trust_remote_code: false
+  max_seq_length: 2048
+  num_epochs: 0
+  learning_rate: 2e-4
+  batch_size: 2
+  gradient_accumulation_steps: 4
+  warmup_steps: 5
+  max_steps: 30
+  save_steps: 30
+  weight_decay: 0.001
+  random_seed: 3407
+  packing: false
+  train_on_completions: true
+  gradient_checkpointing: "unsloth"
+  optim: "adamw_8bit"
+  lr_scheduler_type: "linear"
+
+lora:
+  lora_r: 8
+  lora_alpha: 8
+  lora_dropout: 0.0
+  target_modules:
+    - "all-linear"
+  use_rslora: false
+  use_loftq: false
+  finetune_vision_layers: true
+  finetune_language_layers: true
+  finetune_attention_modules: true
+  finetune_mlp_modules: true
+
+logging:
+  enable_wandb: false
+  wandb_project: "llm-finetuning"
+  enable_tensorboard: false
+  tensorboard_dir: "runs"
+  log_frequency: 10
+
+inference:
+  trust_remote_code: false
+  temperature: 1.0
+  top_p: 0.95
+  top_k: 64
+  min_p: 0.0

studio/backend/assets/configs/model_defaults/gemma/unsloth_gemma-4-E4B-it.yaml

@@ -0,0 +1,47 @@
+# Model defaults for unsloth/gemma-4-E4B-it
+# Also applies to: google/gemma-4-E4B-it, unsloth/gemma-4-E4B-it-GGUF
+
+training:
+  trust_remote_code: false
+  max_seq_length: 2048
+  num_epochs: 0
+  learning_rate: 2e-4
+  batch_size: 2
+  gradient_accumulation_steps: 4
+  warmup_steps: 5
+  max_steps: 30
+  save_steps: 30
+  weight_decay: 0.001
+  random_seed: 3407
+  packing: false
+  train_on_completions: true
+  gradient_checkpointing: "unsloth"
+  optim: "adamw_8bit"
+  lr_scheduler_type: "linear"
+
+lora:
+  lora_r: 8
+  lora_alpha: 8
+  lora_dropout: 0.0
+  target_modules:
+    - "all-linear"
+  use_rslora: false
+  use_loftq: false
+  finetune_vision_layers: true
+  finetune_language_layers: true
+  finetune_attention_modules: true
+  finetune_mlp_modules: true
+
+logging:
+  enable_wandb: false
+  wandb_project: "llm-finetuning"
+  enable_tensorboard: false
+  tensorboard_dir: "runs"
+  log_frequency: 10
+
+inference:
+  trust_remote_code: false
+  temperature: 1.0
+  top_p: 0.95
+  top_k: 64
+  min_p: 0.0

studio/backend/assets/configs/model_defaults/gemma/unsloth_gemma-4-E4B.yaml

@@ -0,0 +1,47 @@
+# Model defaults for unsloth/gemma-4-E4B (base/pretrained)
+# Also applies to: google/gemma-4-E4B
+
+training:
+  trust_remote_code: false
+  max_seq_length: 2048
+  num_epochs: 0
+  learning_rate: 2e-4
+  batch_size: 2
+  gradient_accumulation_steps: 4
+  warmup_steps: 5
+  max_steps: 30
+  save_steps: 30
+  weight_decay: 0.001
+  random_seed: 3407
+  packing: false
+  train_on_completions: true
+  gradient_checkpointing: "unsloth"
+  optim: "adamw_8bit"
+  lr_scheduler_type: "linear"
+
+lora:
+  lora_r: 8
+  lora_alpha: 8
+  lora_dropout: 0.0
+  target_modules:
+    - "all-linear"
+  use_rslora: false
+  use_loftq: false
+  finetune_vision_layers: true
+  finetune_language_layers: true
+  finetune_attention_modules: true
+  finetune_mlp_modules: true
+
+logging:
+  enable_wandb: false
+  wandb_project: "llm-finetuning"
+  enable_tensorboard: false
+  tensorboard_dir: "runs"
+  log_frequency: 10
+
+inference:
+  trust_remote_code: false
+  temperature: 1.0
+  top_p: 0.95
+  top_k: 64
+  min_p: 0.0

studio/backend/utils/models/model_config.py

@@ -159,6 +159,38 @@ MODEL_NAME_MAPPING = {
         "unsloth/gemma-3n-E4B-unsloth-bnb-4bit",
         "google/gemma-3n-E4B",
     ],
+    "unsloth_gemma-4-31B-it.yaml": [
+        "unsloth/gemma-4-31B-it",
+        "google/gemma-4-31B-it",
+    ],
+    "unsloth_gemma-4-26B-A4B-it.yaml": [
+        "unsloth/gemma-4-26B-A4B-it",
+        "google/gemma-4-26B-A4B-it",
+    ],
+    "unsloth_gemma-4-E2B-it.yaml": [
+        "unsloth/gemma-4-E2B-it",
+        "google/gemma-4-E2B-it",
+    ],
+    "unsloth_gemma-4-E4B-it.yaml": [
+        "unsloth/gemma-4-E4B-it",
+        "google/gemma-4-E4B-it",
+    ],
+    "unsloth_gemma-4-31B.yaml": [
+        "unsloth/gemma-4-31B",
+        "google/gemma-4-31B",
+    ],
+    "unsloth_gemma-4-26B-A4B.yaml": [
+        "unsloth/gemma-4-26B-A4B",
+        "google/gemma-4-26B-A4B",
+    ],
+    "unsloth_gemma-4-E2B.yaml": [
+        "unsloth/gemma-4-E2B",
+        "google/gemma-4-E2B",
+    ],
+    "unsloth_gemma-4-E4B.yaml": [
+        "unsloth/gemma-4-E4B",
+        "google/gemma-4-E4B",
+    ],
     "unsloth_gpt-oss-20b.yaml": [
         "openai/gpt-oss-20b",
         "unsloth/gpt-oss-20b-unsloth-bnb-4bit",