chunked cross entropy (#1390)

Co-authored-by: Felipe Mello <felipemello@fb.com>
pytorch · Aug 29, 2024 · 4fba6cd · 4fba6cd
1 parent ec21546
commit 4fba6cd
Show file tree

Hide file tree

Showing 77 changed files with 496 additions and 107 deletions.
diff --git a/docs/source/api_ref_modules.rst b/docs/source/api_ref_modules.rst
@@ -131,3 +131,4 @@ Losses
    rlhf.loss.RSOLoss
    rlhf.loss.IPOLoss
    rlhf.loss.SimPOLoss
+   loss.CEWithChunkedOutputLoss
diff --git a/recipes/configs/code_llama2/7B_full_low_memory.yaml b/recipes/configs/code_llama2/7B_full_low_memory.yaml
@@ -59,7 +59,7 @@ optimizer:
   lr: 2e-5
 optimizer_in_bwd: True
 loss:
-  _component_: torch.nn.CrossEntropyLoss
+  _component_: torchtune.modules.loss.CEWithChunkedOutputLoss
 compile: False
 
 # Training env

diff --git a/recipes/configs/code_llama2/7B_lora_single_device.yaml b/recipes/configs/code_llama2/7B_lora_single_device.yaml
@@ -65,7 +65,7 @@ lr_scheduler:
   _component_: torchtune.modules.get_cosine_schedule_with_warmup
   num_warmup_steps: 100
 loss:
-  _component_: torch.nn.CrossEntropyLoss
+  _component_: torchtune.modules.loss.CEWithChunkedOutputLoss
 compile: False
 
 # Training env

diff --git a/recipes/configs/code_llama2/7B_qlora_single_device.yaml b/recipes/configs/code_llama2/7B_qlora_single_device.yaml
@@ -65,7 +65,7 @@ lr_scheduler:
   _component_: torchtune.modules.get_cosine_schedule_with_warmup
   num_warmup_steps: 100
 loss:
-  _component_: torch.nn.CrossEntropyLoss
+  _component_: torchtune.modules.loss.CEWithChunkedOutputLoss
 compile: False
 
 # Training env

diff --git a/recipes/configs/dev/8B_full_experimental.yaml b/recipes/configs/dev/8B_full_experimental.yaml
@@ -55,7 +55,7 @@ optimizer:
   foreach: False
 
 loss:
-  _component_: torch.nn.CrossEntropyLoss
+  _component_: torchtune.modules.loss.CEWithChunkedOutputLoss
 max_steps_per_epoch: null
 gradient_accumulation_steps: 1
 

diff --git a/recipes/configs/dev/llama2/13B_lora_fsdp2.yaml b/recipes/configs/dev/llama2/13B_lora_fsdp2.yaml
@@ -70,7 +70,7 @@ lr_scheduler:
   num_warmup_steps: 100
 
 loss:
-  _component_: torch.nn.CrossEntropyLoss
+  _component_: torchtune.modules.loss.CEWithChunkedOutputLoss
 
 # Training
 epochs: 1

diff --git a/recipes/configs/dev/llama2/70B_lora_fsdp2.yaml b/recipes/configs/dev/llama2/70B_lora_fsdp2.yaml
@@ -70,7 +70,7 @@ lr_scheduler:
   num_warmup_steps: 100
 
 loss:
-  _component_: torch.nn.CrossEntropyLoss
+  _component_: torchtune.modules.loss.CEWithChunkedOutputLoss
 
 # Training
 epochs: 1

diff --git a/recipes/configs/dev/llama2/70B_qlora_fsdp2.yaml b/recipes/configs/dev/llama2/70B_qlora_fsdp2.yaml
@@ -71,7 +71,7 @@ lr_scheduler:
   num_warmup_steps: 100
 
 loss:
-  _component_: torch.nn.CrossEntropyLoss
+  _component_: torchtune.modules.loss.CEWithChunkedOutputLoss
 
 fsdp:
   cpu_offload: False

diff --git a/recipes/configs/dev/llama2/7B_lora_fsdp2.yaml b/recipes/configs/dev/llama2/7B_lora_fsdp2.yaml
@@ -67,7 +67,7 @@ lr_scheduler:
   num_warmup_steps: 100
 
 loss:
-  _component_: torch.nn.CrossEntropyLoss
+  _component_: torchtune.modules.loss.CEWithChunkedOutputLoss
 
 # Training
 epochs: 1

diff --git a/recipes/configs/dev/llama2/7B_qlora_fsdp2.yaml b/recipes/configs/dev/llama2/7B_qlora_fsdp2.yaml
@@ -67,7 +67,7 @@ lr_scheduler:
   num_warmup_steps: 100
 
 loss:
-  _component_: torch.nn.CrossEntropyLoss
+  _component_: torchtune.modules.loss.CEWithChunkedOutputLoss
 
 fsdp:
   cpu_offload: False

diff --git a/recipes/configs/gemma/2B_full.yaml b/recipes/configs/gemma/2B_full.yaml
@@ -50,7 +50,7 @@ optimizer:
   _component_: torch.optim.AdamW
   lr: 2e-5
 loss:
-  _component_: torch.nn.CrossEntropyLoss
+  _component_: torchtune.modules.loss.CEWithChunkedOutputLoss
 max_steps_per_epoch: null
 gradient_accumulation_steps: 1
 

diff --git a/recipes/configs/gemma/2B_lora.yaml b/recipes/configs/gemma/2B_lora.yaml
@@ -57,7 +57,7 @@ lr_scheduler:
   num_warmup_steps: 100
 
 loss:
-  _component_: torch.nn.CrossEntropyLoss
+  _component_: torchtune.modules.loss.CEWithChunkedOutputLoss
 
 # Fine-tuning arguments
 batch_size: 4

diff --git a/recipes/configs/gemma/2B_lora_single_device.yaml b/recipes/configs/gemma/2B_lora_single_device.yaml
@@ -56,7 +56,7 @@ lr_scheduler:
   num_warmup_steps: 100
 
 loss:
-  _component_: torch.nn.CrossEntropyLoss
+  _component_: torchtune.modules.loss.CEWithChunkedOutputLoss
 
 # Fine-tuning arguments
 batch_size: 4

diff --git a/recipes/configs/gemma/2B_qlora_single_device.yaml b/recipes/configs/gemma/2B_qlora_single_device.yaml
@@ -56,7 +56,7 @@ lr_scheduler:
   num_warmup_steps: 100
 
 loss:
-  _component_: torch.nn.CrossEntropyLoss
+  _component_: torchtune.modules.loss.CEWithChunkedOutputLoss
 
 # Fine-tuning arguments
 batch_size: 4

diff --git a/recipes/configs/gemma/7B_full.yaml b/recipes/configs/gemma/7B_full.yaml
@@ -52,7 +52,7 @@ optimizer:
   _component_: torch.optim.AdamW
   lr: 2e-5
 loss:
-  _component_: torch.nn.CrossEntropyLoss
+  _component_: torchtune.modules.loss.CEWithChunkedOutputLoss
 max_steps_per_epoch: null
 gradient_accumulation_steps: 1
 

diff --git a/recipes/configs/gemma/7B_lora.yaml b/recipes/configs/gemma/7B_lora.yaml
@@ -59,7 +59,7 @@ lr_scheduler:
   num_warmup_steps: 100
 
 loss:
-  _component_: torch.nn.CrossEntropyLoss
+  _component_: torchtune.modules.loss.CEWithChunkedOutputLoss
 
 # Fine-tuning arguments
 batch_size: 4

diff --git a/recipes/configs/gemma/7B_lora_single_device.yaml b/recipes/configs/gemma/7B_lora_single_device.yaml
@@ -58,7 +58,7 @@ lr_scheduler:
   num_warmup_steps: 100
 
 loss:
-  _component_: torch.nn.CrossEntropyLoss
+  _component_: torchtune.modules.loss.CEWithChunkedOutputLoss
 
 # Fine-tuning arguments
 batch_size: 8

diff --git a/recipes/configs/gemma/7B_qlora_single_device.yaml b/recipes/configs/gemma/7B_qlora_single_device.yaml
@@ -58,7 +58,7 @@ lr_scheduler:
   num_warmup_steps: 100
 
 loss:
-  _component_: torch.nn.CrossEntropyLoss
+  _component_: torchtune.modules.loss.CEWithChunkedOutputLoss
 
 # Fine-tuning arguments
 batch_size: 4

diff --git a/recipes/configs/llama2/13B_full.yaml b/recipes/configs/llama2/13B_full.yaml
@@ -54,7 +54,7 @@ optimizer:
   _component_: torch.optim.AdamW
   lr: 2e-5
 loss:
-  _component_: torch.nn.CrossEntropyLoss
+  _component_: torchtune.modules.loss.CEWithChunkedOutputLoss
 max_steps_per_epoch: null
 gradient_accumulation_steps: 1
 

diff --git a/recipes/configs/llama2/13B_lora.yaml b/recipes/configs/llama2/13B_lora.yaml
@@ -66,7 +66,7 @@ lr_scheduler:
   num_warmup_steps: 100
 
 loss:
-  _component_: torch.nn.CrossEntropyLoss
+  _component_: torchtune.modules.loss.CEWithChunkedOutputLoss
 
 # Training
 epochs: 1

diff --git a/recipes/configs/llama2/13B_qlora_single_device.yaml b/recipes/configs/llama2/13B_qlora_single_device.yaml
@@ -61,7 +61,7 @@ lr_scheduler:
   num_warmup_steps: 100
 
 loss:
-  _component_: torch.nn.CrossEntropyLoss
+  _component_: torchtune.modules.loss.CEWithChunkedOutputLoss
 
 # Training
 epochs: 1

diff --git a/recipes/configs/llama2/70B_lora.yaml b/recipes/configs/llama2/70B_lora.yaml
@@ -66,7 +66,7 @@ lr_scheduler:
   num_warmup_steps: 100
 
 loss:
-  _component_: torch.nn.CrossEntropyLoss
+  _component_: torchtune.modules.loss.CEWithChunkedOutputLoss
 
 # Training
 epochs: 1

diff --git a/recipes/configs/llama2/7B_full.yaml b/recipes/configs/llama2/7B_full.yaml
@@ -53,7 +53,7 @@ optimizer:
   _component_: torch.optim.AdamW
   lr: 2e-5
 loss:
-  _component_: torch.nn.CrossEntropyLoss
+  _component_: torchtune.modules.loss.CEWithChunkedOutputLoss
 max_steps_per_epoch: null
 gradient_accumulation_steps: 1
 

diff --git a/recipes/configs/llama2/7B_full_low_memory.yaml b/recipes/configs/llama2/7B_full_low_memory.yaml
@@ -56,7 +56,7 @@ optimizer:
   lr: 2e-5
 optimizer_in_bwd: True
 loss:
-  _component_: torch.nn.CrossEntropyLoss
+  _component_: torchtune.modules.loss.CEWithChunkedOutputLoss
 max_steps_per_epoch: null
 gradient_accumulation_steps: 1
 compile: False

diff --git a/recipes/configs/llama2/7B_lora.yaml b/recipes/configs/llama2/7B_lora.yaml
@@ -63,7 +63,7 @@ lr_scheduler:
   num_warmup_steps: 100
 
 loss:
-  _component_: torch.nn.CrossEntropyLoss
+  _component_: torchtune.modules.loss.CEWithChunkedOutputLoss
 
 # Training
 epochs: 1

diff --git a/recipes/configs/llama2/7B_lora_single_device.yaml b/recipes/configs/llama2/7B_lora_single_device.yaml
@@ -61,7 +61,7 @@ lr_scheduler:
   num_warmup_steps: 100
 
 loss:
-  _component_: torch.nn.CrossEntropyLoss
+  _component_: torchtune.modules.loss.CEWithChunkedOutputLoss
 
 # Training
 epochs: 1

diff --git a/recipes/configs/llama2/7B_qat_full.yaml b/recipes/configs/llama2/7B_qat_full.yaml
@@ -49,7 +49,7 @@ optimizer:
   _component_: torch.optim.AdamW
   lr: 2e-5
 loss:
-  _component_: torch.nn.CrossEntropyLoss
+  _component_: torchtune.modules.loss.CEWithChunkedOutputLoss
 max_steps_per_epoch: null
 gradient_accumulation_steps: 1
 

diff --git a/recipes/configs/llama2/7B_qlora_single_device.yaml b/recipes/configs/llama2/7B_qlora_single_device.yaml
@@ -60,7 +60,7 @@ lr_scheduler:
   num_warmup_steps: 100
 
 loss:
-  _component_: torch.nn.CrossEntropyLoss
+  _component_: torchtune.modules.loss.CEWithChunkedOutputLoss
 
 # Training
 epochs: 1

diff --git a/recipes/configs/llama3/70B_full.yaml b/recipes/configs/llama3/70B_full.yaml
@@ -85,7 +85,7 @@ optimizer:
   fused: True
 
 loss:
-  _component_: torch.nn.CrossEntropyLoss
+  _component_: torchtune.modules.loss.CEWithChunkedOutputLoss
 max_steps_per_epoch: null
 gradient_accumulation_steps: 1
 

diff --git a/recipes/configs/llama3/70B_lora.yaml b/recipes/configs/llama3/70B_lora.yaml
@@ -81,7 +81,7 @@ lr_scheduler:
   num_warmup_steps: 100
 
 loss:
-  _component_: torch.nn.CrossEntropyLoss
+  _component_: torchtune.modules.loss.CEWithChunkedOutputLoss
 
 # Training
 epochs: 1

diff --git a/recipes/configs/llama3/8B_full.yaml b/recipes/configs/llama3/8B_full.yaml
@@ -55,7 +55,7 @@ optimizer:
   foreach: False
 
 loss:
-  _component_: torch.nn.CrossEntropyLoss
+  _component_: torchtune.modules.loss.CEWithChunkedOutputLoss
 max_steps_per_epoch: null
 gradient_accumulation_steps: 1
 

diff --git a/recipes/configs/llama3/8B_full_single_device.yaml b/recipes/configs/llama3/8B_full_single_device.yaml
@@ -54,7 +54,7 @@ optimizer:
   _component_: bitsandbytes.optim.PagedAdamW8bit
   lr: 2e-5
 loss:
-  _component_: torch.nn.CrossEntropyLoss
+  _component_: torchtune.modules.loss.CEWithChunkedOutputLoss
 max_steps_per_epoch: null
 gradient_accumulation_steps: 1
 optimizer_in_bwd: True

diff --git a/recipes/configs/llama3/8B_lora.yaml b/recipes/configs/llama3/8B_lora.yaml
@@ -61,7 +61,7 @@ lr_scheduler:
   num_warmup_steps: 100
 
 loss:
-  _component_: torch.nn.CrossEntropyLoss
+  _component_: torchtune.modules.loss.CEWithChunkedOutputLoss
 
 # Training
 epochs: 1

diff --git a/recipes/configs/llama3/8B_lora_single_device.yaml b/recipes/configs/llama3/8B_lora_single_device.yaml
@@ -60,7 +60,7 @@ lr_scheduler:
   num_warmup_steps: 100
 
 loss:
-  _component_: torch.nn.CrossEntropyLoss
+  _component_: torchtune.modules.loss.CEWithChunkedOutputLoss
 
 # Training
 epochs: 1

diff --git a/recipes/configs/llama3/8B_qat_full.yaml b/recipes/configs/llama3/8B_qat_full.yaml
@@ -55,7 +55,7 @@ optimizer:
   foreach: False
 
 loss:
-  _component_: torch.nn.CrossEntropyLoss
+  _component_: torchtune.modules.loss.CEWithChunkedOutputLoss
 max_steps_per_epoch: null
 gradient_accumulation_steps: 1
 

diff --git a/recipes/configs/llama3/8B_qlora_single_device.yaml b/recipes/configs/llama3/8B_qlora_single_device.yaml
@@ -59,7 +59,7 @@ lr_scheduler:
   num_warmup_steps: 100
 
 loss:
-  _component_: torch.nn.CrossEntropyLoss
+  _component_: torchtune.modules.loss.CEWithChunkedOutputLoss
 
 # Training
 epochs: 1

diff --git a/recipes/configs/llama3_1/70B_full.yaml b/recipes/configs/llama3_1/70B_full.yaml
@@ -85,7 +85,7 @@ optimizer:
   fused: True
 
 loss:
-  _component_: torch.nn.CrossEntropyLoss
+  _component_: torchtune.modules.loss.CEWithChunkedOutputLoss
 max_steps_per_epoch: null
 gradient_accumulation_steps: 1
 

diff --git a/recipes/configs/llama3_1/70B_lora.yaml b/recipes/configs/llama3_1/70B_lora.yaml
@@ -80,7 +80,7 @@ lr_scheduler:
   num_warmup_steps: 100
 
 loss:
-  _component_: torch.nn.CrossEntropyLoss
+  _component_: torchtune.modules.loss.CEWithChunkedOutputLoss
 
 # Training
 epochs: 1

diff --git a/recipes/configs/llama3_1/8B_full.yaml b/recipes/configs/llama3_1/8B_full.yaml
@@ -58,7 +58,7 @@ optimizer:
   foreach: False
 
 loss:
-  _component_: torch.nn.CrossEntropyLoss
+  _component_: torchtune.modules.loss.CEWithChunkedOutputLoss
 max_steps_per_epoch: null
 gradient_accumulation_steps: 1
 

diff --git a/recipes/configs/llama3_1/8B_full_single_device.yaml b/recipes/configs/llama3_1/8B_full_single_device.yaml
@@ -57,7 +57,7 @@ optimizer:
   _component_: bitsandbytes.optim.PagedAdamW8bit
   lr: 2e-5
 loss:
-  _component_: torch.nn.CrossEntropyLoss
+  _component_: torchtune.modules.loss.CEWithChunkedOutputLoss
 max_steps_per_epoch: null
 gradient_accumulation_steps: 1
 optimizer_in_bwd: True
@@ -79,3 +79,28 @@ metric_logger:
 output_dir: /tmp/full-llama3.1-finetune
 log_every_n_steps: 1
 log_peak_memory_stats: False
+
+# Profiler (disabled)
+profiler:
+  _component_: torchtune.utils.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: True
+  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: 1
+  warmup_steps: 2
+  active_steps: 1
+  num_cycles: 1
diff --git a/recipes/configs/llama3_1/8B_lora.yaml b/recipes/configs/llama3_1/8B_lora.yaml
@@ -64,7 +64,7 @@ lr_scheduler:
   num_warmup_steps: 100
 
 loss:
-  _component_: torch.nn.CrossEntropyLoss
+  _component_: torchtune.modules.loss.CEWithChunkedOutputLoss
 
 # Training
 epochs: 1