removing deepspeed guard for LoRA Triton kernels

* fix(example): align example to correct adapter

* fix: add missing load in 4 bit

examples/gemma3/gemma-3-4b-qlora.yml

@@ -1,6 +1,8 @@
 base_model: google/gemma-3-4b-it
 strict: false
 
+load_in_4bit: true
+
 # gemma3 doesn't seem to play nice with ddp
 ddp_find_unused_parameters: true
 
@@ -17,7 +19,7 @@ dataset_prepared_path: last_run_prepared
 val_set_size: 0.01
 output_dir: ./outputs/out
 
-adapter: lora
+adapter: qlora
 lora_model_dir:
 
 sequence_len: 2048

examples/gemma3/gemma-3-4b-vision-qlora.yml

@@ -2,6 +2,8 @@ base_model: google/gemma-3-4b-it
 processor_type: AutoProcessor
 strict: false
 
+load_in_4bit: true
+
 # these 3 lines are needed for now to handle vision chat templates w images
 skip_prepare_dataset: true
 remove_unused_columns: false

src/axolotl/utils/config/__init__.py

@@ -78,6 +78,7 @@ def resolve_dtype(cfg):
         cfg.bf16 = False
     else:
         torch.backends.cuda.matmul.allow_tf32 = cfg.tf32 or False
+        torch.backends.cudnn.allow_tf32 = cfg.tf32 or False
         if cfg.bf16:
             cfg.fp16 = False
 

src/axolotl/utils/schemas/config.py

@@ -1224,17 +1224,12 @@ class AxolotlConfigWCapabilities(AxolotlInputConfig):
         ):
             capabilities = data.get("capabilities")
             is_fsdp = data.get("fsdp") is not None
-            is_deepspeed = data.get("deepspeed") is not None
 
             if capabilities and capabilities.get("n_gpu", 0) > 1:
                 if is_fsdp:
                     raise ValueError(
                         "lora_mlp_kernel, lora_qkv_kernel, and lora_o_kernel are not compatible with FSDP."
                     )
-                if is_deepspeed:
-                    raise ValueError(
-                        "lora_mlp_kernel, lora_qkv_kernel, and lora_o_kernel are not compatible with DeepSpeed."
-                    )
         return data
 
     @model_validator(mode="before")