Solves #2598

recipes/configs/qwen2_5/14B_to_7B_KD_lora_single_device.yaml

@@ -0,0 +1,136 @@
+# Config for single device knowledge distillation (KD) in knowledge_distillation_single_device.py
+# using a Qwen14B teacher and student model
+#
+# This config assumes that you've ran the following commands before launching KD:
+# First download the student and teacher models
+#   tune download Qwen/Qwen2.5-14B-Instruct
+#   tune download Qwen/Qwen2.5-7B-Instruct
+#
+# You get better results using KD if the teacher model has already been fine-tuned on the target dataset:
+#   tune run lora_finetune_single_device --config qwen2_5/14B_lora_single_device.yaml
+#
+# To launch on a single device, run the following command from root:
+#   tune run knowledge_distillation_single_device --config qwen2_5/14B_to_7B_KD_lora_single_device.yaml
+#
+# This config works only for training on single device.
+
+
+output_dir: /tmp/kd_7B # /tmp may be deleted by your system. Change it to your preference.
+
+# Model Arguments
+model:
+  _component_: torchtune.models.qwen2_5.lora_qwen2_5_7b
+  lora_attn_modules: ['q_proj', 'v_proj', 'output_proj']
+  apply_lora_to_mlp: True
+  apply_lora_to_output: False
+  lora_rank: 64  # higher increases accuracy and memory
+  lora_alpha: 128  # usually alpha=2*rank
+  lora_dropout: 0.0
+
+teacher_model:
+  _component_: torchtune.models.qwen2_5.lora_qwen2_5_14b_instruct
+  lora_attn_modules: ['q_proj', 'v_proj', 'output_proj']
+
+# Tokenizer
+tokenizer:
+  _component_: torchtune.models.qwen2_5.qwen2_5_tokenizer
+  path: /Qwen2.5-7B-Instruct/vocab.json
+  merges_file: /Qwen2.5-7B-Instruct/merges.txt
+  max_seq_len: null
+
+checkpointer:
+  _component_: torchtune.training.FullModelHFCheckpointer
+  checkpoint_dir: /tmp/Qwen2.5-7B-Instruct
+  checkpoint_files:
+    filename_format: model-{}-of-{}.safetensors
+    max_filename: "00050"
+  recipe_checkpoint: null
+  output_dir: ${output_dir}
+  model_type: QWEN2
+resume_from_checkpoint: False
+save_adapter_weights_only: False
+
+# Teacher checkpoint
+teacher_checkpointer:
+  _component_: torchtune.training.FullModelHFCheckpointer
+  checkpoint_dir: /tmp/Qwen2.5-14B-Instruct
+  checkpoint_files:
+    filename_format: model-{}-of-{}.safetensors
+    max_filename: "00050"
+  recipe_checkpoint: null
+  output_dir: ${output_dir}
+  model_type: QWEN2
+
+# Dataset and Sampler
+dataset:
+  _component_: torchtune.datasets.alpaca_cleaned_dataset
+  packed: False  # True increases speed
+seed: null
+shuffle: True
+batch_size: 2
+
+# Optimizer and Scheduler
+optimizer:
+  _component_: torch.optim.AdamW
+  fused: True
+  weight_decay: 0.01
+  lr: 3e-4
+lr_scheduler:
+  _component_: torchtune.training.lr_schedulers.get_cosine_schedule_with_warmup
+  num_warmup_steps: 100
+
+loss:
+  _component_: torchtune.modules.loss.CEWithChunkedOutputLoss
+
+kd_loss:
+  _component_: torchtune.modules.loss.ForwardKLWithChunkedOutputLoss
+kd_ratio: 0.5
+
+# Training
+epochs: 5
+max_steps_per_epoch: null
+gradient_accumulation_steps: 8  # Use to increase effective batch size
+clip_grad_norm: null
+compile: False  # torch.compile the model + loss, True increases speed + decreases memory
+
+# Logging
+metric_logger:
+  _component_: torchtune.training.metric_logging.DiskLogger
+  log_dir: ${output_dir}/logs
+log_every_n_steps: 2
+log_peak_memory_stats: True
+
+# Environment
+device: cuda
+dtype: bf16
+
+# Activations Memory
+enable_activation_checkpointing: True  # True reduces memory
+enable_activation_offloading: False  # True reduces memory
+
+
+
+# Profiler (disabled)
+profiler:
+  _component_: torchtune.training.setup_torch_profiler
+  enabled: False
+
+  #Output directory of trace artifacts
+  output_dir: ${output_dir}/profiling_outputs
+
+  #`torch.profiler.ProfilerActivity` types to trace
+  cpu: True
+  cuda: True
+
+  #trace options passed to `torch.profiler.profile`
+  profile_memory: False
+  with_stack: False
+  record_shapes: True
+  with_flops: False
+
+  # `torch.profiler.schedule` options:
+  # wait_steps -> wait, warmup_steps -> warmup, active_steps -> active, num_cycles -> repeat
+  wait_steps: 5
+  warmup_steps: 3
+  active_steps: 2
+  num_cycles: 1

torchtune/modules/loss/kd_losses.py

@@ -25,6 +25,21 @@ def __init__(self, ignore_index: int = -100):
         super().__init__()
         self.ignore_index = ignore_index
 
+    def pad_logits(self,student_logits, teacher_logits):
+        student_size, teacher_size = student_logits.size(-1), teacher_logits.size(-1)
+        if student_size != teacher_size:
+            pad_size = abs(student_size - teacher_size)
+            pad_tensor = torch.zeros(
+                (*teacher_logits.shape[:-1], pad_size),
+                dtype=teacher_logits.dtype,
+                device=teacher_logits.device,
+            )
+            return (
+                (torch.cat([student_logits, pad_tensor], dim=-1), teacher_logits)
+                if student_size < teacher_size
+                else (student_logits, torch.cat([teacher_logits, pad_tensor], dim=-1))
+            )
+        return student_logits, teacher_logits
     def forward(
         self,
         student_logits: torch.Tensor,
@@ -49,6 +64,16 @@ def forward(
         teacher_prob = F.softmax(teacher_logits, dim=-1, dtype=torch.float32)
         inf_mask = torch.isinf(student_logits)
         student_logprob = F.log_softmax(student_logits, dim=-1, dtype=torch.float32)
+        student_logprob, teacher_prob = self.pad_logits(
+            student_logprob,
+            teacher_prob,
+        )
+        inf_mask,_ = self.pad_logits(
+            inf_mask,
+            teacher_prob,
+        )
+        inf_mask = torch.isinf(student_logprob)
+
         prod_probs = torch.masked_fill(teacher_prob * student_logprob, inf_mask, 0)
         x = torch.sum(prod_probs, dim=-1).view(-1)
         mask = (labels != self.ignore_index).int()
@@ -61,6 +86,7 @@ def forward(
         return -torch.sum(x * mask.view(-1), dim=0) / torch.sum(mask.view(-1), dim=0)
 
 
+
 class ReverseKLLoss(torch.nn.Module):
     """
     The Kullback-Leibler divergence loss for valid indexes.