fix: Added sequence packing keys to SFT and GRPO recipes (#805)

ahmadki · web-flow · commit fa21d3c866a8 · 2025-08-04T21:42:51.000Z
Signed-off-by: Ahmad Kiswani &lt;kiswani.ahmad@gmail.com&gt;
diff --git a/examples/configs/recipes/llm/grpo-gemma3-1b-it-1n8g-fsdp2tp1.yaml b/examples/configs/recipes/llm/grpo-gemma3-1b-it-1n8g-fsdp2tp1.yaml
@@ -49,6 +49,10 @@ policy:
     sequence_length_round: 64
   sequence_packing:
     enabled: false
+    train_mb_tokens: ${mul:${policy.max_total_sequence_length}, ${policy.train_micro_batch_size}}
+    logprob_mb_tokens: ${mul:${policy.max_total_sequence_length}, ${policy.logprob_batch_size}}
+    algorithm: "modified_first_fit_decreasing"
+    sequence_length_round: 64
   make_sequence_length_divisible_by: 1
   max_grad_norm: 1
   optimizer:
diff --git a/examples/configs/recipes/llm/grpo-gemma3-27b-it-16n8g-fsdp2tp8sp-actckpt-long.yaml b/examples/configs/recipes/llm/grpo-gemma3-27b-it-16n8g-fsdp2tp8sp-actckpt-long.yaml
@@ -50,6 +50,10 @@ policy:
     sequence_length_round: 64
   sequence_packing:
     enabled: false
+    train_mb_tokens: ${mul:${policy.max_total_sequence_length}, ${policy.train_micro_batch_size}}
+    logprob_mb_tokens: ${mul:${policy.max_total_sequence_length}, ${policy.logprob_batch_size}}
+    algorithm: "modified_first_fit_decreasing"
+    sequence_length_round: 64
   make_sequence_length_divisible_by: 8
   max_grad_norm: 1
   optimizer:
diff --git a/examples/configs/recipes/llm/grpo-llama3.1-8b-instruct-4n8g-fsdp2tp1-long.v3.yaml b/examples/configs/recipes/llm/grpo-llama3.1-8b-instruct-4n8g-fsdp2tp1-long.v3.yaml
@@ -49,6 +49,10 @@ policy:
     sequence_length_round: 64
   sequence_packing:
     enabled: false
+    train_mb_tokens: ${mul:${policy.max_total_sequence_length}, ${policy.train_micro_batch_size}}
+    logprob_mb_tokens: ${mul:${policy.max_total_sequence_length}, ${policy.logprob_batch_size}}
+    algorithm: "modified_first_fit_decreasing"
+    sequence_length_round: 64
   make_sequence_length_divisible_by: 1
   max_grad_norm: 1
   optimizer:
diff --git a/examples/configs/recipes/llm/grpo-llama3.2-1b-instruct-1n8g-fsdp2tp1.v3.yaml b/examples/configs/recipes/llm/grpo-llama3.2-1b-instruct-1n8g-fsdp2tp1.v3.yaml
@@ -49,6 +49,10 @@ policy:
     sequence_length_round: 64
   sequence_packing:
     enabled: false
+    train_mb_tokens: ${mul:${policy.max_total_sequence_length}, ${policy.train_micro_batch_size}}
+    logprob_mb_tokens: ${mul:${policy.max_total_sequence_length}, ${policy.logprob_batch_size}}
+    algorithm: "modified_first_fit_decreasing"
+    sequence_length_round: 64
   make_sequence_length_divisible_by: 1
   max_grad_norm: 1
   optimizer:
diff --git a/examples/configs/recipes/llm/grpo-qwen2.5-32b-32n8g-fsdp2tp8sp-actckpt-long.v3.yaml b/examples/configs/recipes/llm/grpo-qwen2.5-32b-32n8g-fsdp2tp8sp-actckpt-long.v3.yaml
@@ -49,6 +49,10 @@ policy:
     sequence_length_round: 64
   sequence_packing:
     enabled: false
+    train_mb_tokens: ${mul:${policy.max_total_sequence_length}, ${policy.train_micro_batch_size}}
+    logprob_mb_tokens: ${mul:${policy.max_total_sequence_length}, ${policy.logprob_batch_size}}
+    algorithm: "modified_first_fit_decreasing"
+    sequence_length_round: 64
   make_sequence_length_divisible_by: 8
   max_grad_norm: 1
   optimizer:
diff --git a/examples/configs/recipes/llm/grpo-qwen2.5-32b-32n8g-fsdp2tp8sp-actckpt.v3.yaml b/examples/configs/recipes/llm/grpo-qwen2.5-32b-32n8g-fsdp2tp8sp-actckpt.v3.yaml
@@ -49,6 +49,10 @@ policy:
     sequence_length_round: 64
   sequence_packing:
     enabled: false
+    train_mb_tokens: ${mul:${policy.max_total_sequence_length}, ${policy.train_micro_batch_size}}
+    logprob_mb_tokens: ${mul:${policy.max_total_sequence_length}, ${policy.logprob_batch_size}}
+    algorithm: "modified_first_fit_decreasing"
+    sequence_length_round: 64
   make_sequence_length_divisible_by: 8
   max_grad_norm: 1
   optimizer:
diff --git a/examples/configs/recipes/llm/grpo-qwen2.5-7b-instruct-4n8g-fsdp2tp4sp.v3.yaml b/examples/configs/recipes/llm/grpo-qwen2.5-7b-instruct-4n8g-fsdp2tp4sp.v3.yaml
@@ -49,6 +49,10 @@ policy:
     sequence_length_round: 64
   sequence_packing:
     enabled: false
+    train_mb_tokens: ${mul:${policy.max_total_sequence_length}, ${policy.train_micro_batch_size}}
+    logprob_mb_tokens: ${mul:${policy.max_total_sequence_length}, ${policy.logprob_batch_size}}
+    algorithm: "modified_first_fit_decreasing"
+    sequence_length_round: 64
   make_sequence_length_divisible_by: 4
   max_grad_norm: 1
   optimizer:
diff --git a/examples/configs/recipes/llm/grpo-qwen2.5-math-1.5b-instruct-1n8g-fsdp2tp1.v3.yaml b/examples/configs/recipes/llm/grpo-qwen2.5-math-1.5b-instruct-1n8g-fsdp2tp1.v3.yaml
@@ -49,6 +49,10 @@ policy:
     sequence_length_round: 64
   sequence_packing:
     enabled: false
+    train_mb_tokens: ${mul:${policy.max_total_sequence_length}, ${policy.train_micro_batch_size}}
+    logprob_mb_tokens: ${mul:${policy.max_total_sequence_length}, ${policy.logprob_batch_size}}
+    algorithm: "modified_first_fit_decreasing"
+    sequence_length_round: 64
   make_sequence_length_divisible_by: 1
   max_grad_norm: 1
   optimizer:
diff --git a/examples/configs/recipes/llm/sft-llama3.1-8b-instruct-1n8g-fsdp2tp1-long.v2.yaml b/examples/configs/recipes/llm/sft-llama3.1-8b-instruct-1n8g-fsdp2tp1-long.v2.yaml
@@ -35,6 +35,9 @@ policy:
     enabled: false
   sequence_packing:
     enabled: false
+    train_mb_tokens: ${mul:${policy.max_total_sequence_length}, ${policy.train_micro_batch_size}}
+    algorithm: "modified_first_fit_decreasing"
+    sequence_length_round: 64
   make_sequence_length_divisible_by: 1
   max_grad_norm: 1
   optimizer:
diff --git a/examples/configs/recipes/llm/sft-llama3.1-8b-instruct-1n8g-fsdp2tp2sp.v2.yaml b/examples/configs/recipes/llm/sft-llama3.1-8b-instruct-1n8g-fsdp2tp2sp.v2.yaml
@@ -35,6 +35,9 @@ policy:
     enabled: false
   sequence_packing:
     enabled: false
+    train_mb_tokens: ${mul:${policy.max_total_sequence_length}, ${policy.train_micro_batch_size}}
+    algorithm: "modified_first_fit_decreasing"
+    sequence_length_round: 64
   make_sequence_length_divisible_by: 2
   max_grad_norm: 1
   optimizer:
diff --git a/examples/configs/recipes/llm/sft-llama3.1-8b-instruct-1n8g-megatron.yaml b/examples/configs/recipes/llm/sft-llama3.1-8b-instruct-1n8g-megatron.yaml
@@ -29,6 +29,9 @@ policy:
     enabled: false
   sequence_packing:
     enabled: false
+    train_mb_tokens: ${mul:${policy.max_total_sequence_length}, ${policy.train_micro_batch_size}}
+    algorithm: "modified_first_fit_decreasing"
+    sequence_length_round: 64
   make_sequence_length_divisible_by: ${policy.megatron_cfg.tensor_model_parallel_size}
   max_grad_norm: 1
   optimizer: null
diff --git a/examples/configs/recipes/llm/sft-llama3.2-1b-1n8g-fsdp2tp1.v2.yaml b/examples/configs/recipes/llm/sft-llama3.2-1b-1n8g-fsdp2tp1.v2.yaml
@@ -35,6 +35,9 @@ policy:
     enabled: false
   sequence_packing:
     enabled: false
+    train_mb_tokens: ${mul:${policy.max_total_sequence_length}, ${policy.train_micro_batch_size}}
+    algorithm: "modified_first_fit_decreasing"
+    sequence_length_round: 64
   make_sequence_length_divisible_by: 1
   max_grad_norm: 1
   optimizer:
diff --git a/examples/configs/recipes/llm/sft-qwen2.5-32b-4n8g-fsdp2tp8sp-actckpt.v2.yaml b/examples/configs/recipes/llm/sft-qwen2.5-32b-4n8g-fsdp2tp8sp-actckpt.v2.yaml
@@ -35,6 +35,9 @@ policy:
     enabled: false
   sequence_packing:
     enabled: false
+    train_mb_tokens: ${mul:${policy.max_total_sequence_length}, ${policy.train_micro_batch_size}}
+    algorithm: "modified_first_fit_decreasing"
+    sequence_length_round: 64
   make_sequence_length_divisible_by: 8
   max_grad_norm: 1
   optimizer:
