Merge branch 'core-pipeline-fix' into 'main'

Fix bug in core pipeline schedule

See merge request ADLR/megatron-lm!552

.gitlab-ci.yml

@@ -87,6 +87,7 @@ unit_tests:
     - export BUILD_DIR=`pwd`
     - export RUN_NAME=${RUN_MODEL}_tp${TP_SIZE}_pp${PP_SIZE}_${NUM_NODES}nodes_${MAX_STEPS}steps
     - export TP_SIZE PP_SIZE NUM_NODES MAX_STEPS VP_SIZE
+    - export MBS GBS
     - export DATA_DIR=$DATA_DIR
     - echo "Run name is $RUN_NAME"
     - mkdir -p $SELENE_ADLR_CI_PATH/$CI_PIPELINE_ID/$RUN_NAME/checkpoints
@@ -100,7 +101,7 @@ unit_tests:
     - export RESULTS_DIR=$BASE_DIR/results
     - export CHECKPOINTS_DIR=$BASE_DIR/checkpoints
     - echo "Submitting job"
-    - sbatch_submission=`sbatch $BUILD_DIR/tests/functional_tests/test_scripts/$RUN_MODEL/sbatch_${RUN_MODEL}_distributed_test.sh --export=BASE_DIR,BUILD_DIR,DATA_DIR,TP_SIZE,PP_SIZE,NUM_NODES,MAX_STEPS,VP_SIZE`
+    - sbatch_submission=`sbatch $BUILD_DIR/tests/functional_tests/test_scripts/$RUN_MODEL/sbatch_${RUN_MODEL}_distributed_test.sh --export=BASE_DIR,BUILD_DIR,DATA_DIR,TP_SIZE,PP_SIZE,NUM_NODES,MAX_STEPS,VP_SIZE,MBS,GBS`
     - export SLURM_JOBID=$(echo $sbatch_submission| grep 'Submitted batch job' | awk '{ print $4 }');
     - bash $BUILD_DIR/tests/functional_tests/shell_test_utils/jobwait.sh $SLURM_JOBID
     - \[ ! -z ${SLURM_JOBID} \] && echo -e " --------------------------------------------------\n"
@@ -167,6 +168,19 @@ train.gpt3.345m_tp1_pp2_1node_50steps:
     TIME_LIMIT: "20:00"
     TEST_LEVEL: L0 
 
+train.gpt3.345m_tp1_pp4_1node_50steps:
+  <<: *selene-test-launcher
+  variables:
+    <<: [*VARS]
+    RUN_MODEL: gpt3
+    TP_SIZE: 1
+    PP_SIZE: 4
+    VP_SIZE: 1
+    NUM_NODES: 1
+    MAX_STEPS: 50
+    TIME_LIMIT: "20:00"
+    TEST_LEVEL: L0
+
 resume.checkpoint.gpt3.345m_tp1_pp2_1node:
   <<: *selene-test-resume-checkpoint-launcher
   variables:

megatron/core/pipeline_parallel/schedules.py

@@ -593,7 +593,7 @@ def forward_backward_pipelining_with_interleaving(*,
     if not forward_only:
         if all_warmup_microbatches:
             output_tensor_grads[num_model_chunks-1].append(
-                p2p_communication.recv_backward(tensor_shape, timers=timers))
+                p2p_communication.recv_backward(tensor_shape, dtype=dtype, timers=timers))
         for k in range(num_microbatches_remaining, total_num_microbatches):
             input_tensor_grad = backward_step_helper(k)
             next_backward_model_chunk_id = get_model_chunk_id(k+1, forward=False)

tests/functional_tests/python_test_utils/test_ci_pipeline.py

@@ -53,9 +53,11 @@ class TestCIPipeline:
             raise FileNotFoundError("Expected data is none")
         expected = self.expected[loss_type]
         expected_list = expected["values"]
+        print(expected_list)
         actual_list = read_tb_logs_as_list(LOGS_DIR, loss_type)
         assert actual_list is not None, f"No TensorBoard events file was found in the logs for {loss_type}."
         for i, step in enumerate(range(expected["start_step"], expected["end_step"], expected["step_interval"])):
+            print(f"Checking step {step} against expected {i}")
             if test_type == TypeOfTest.APPROX:
                 assert actual_list[step] == pytest.approx(expected=expected_list[i], rel=self.margin_loss), f"{self.job_name} : The loss at step {step} should be approximately {expected_list[i]} but it is {actual_list[step]}."
             else:

tests/functional_tests/test_results/gpt3/gpt3_tp1_pp4_1nodes_50steps.json

+{"lm loss": {"start_step": 0, "end_step": 45, "step_interval": 5, "values": [10.7947, 10.85294, 10.87058, 10.83388, 10.83012, 10.78726, 10.56378, 10.57311, 10.48692]}, "num-zeros": {"start_step": 0, "end_step": 29, "step_interval": 5, "values": [2452.0, 2818.0, 2036.0, 2662.0, 2651.0, 2422.0]}, "iteration_timing_avg": 0.1187023333333333}

tests/functional_tests/test_scripts/gpt3/pretrain_gpt3_distributed_test.sh

@@ -7,7 +7,9 @@ TP_SIZE=$4
 PP_SIZE=$5
 NNODES=$6
 MAX_STEPS=$7
-
+VP_SIZE=$8
+MBS=$9
+GBS=${10}
 GPUS_PER_NODE=8
 # Change for multinode config
 MASTER_ADDR=localhost
@@ -30,8 +32,8 @@ python -m torch.distributed.launch $DISTRIBUTED_ARGS \
        --log-validation-ppl-to-tensorboard \
        --log-timers-to-tensorboard \
        --tensorboard-dir ${TENSORBOARD_DIR} \
-       --micro-batch-size 4 \
-       --global-batch-size 32 \
+       --micro-batch-size ${MBS:-4} \
+       --global-batch-size ${GBS:-32} \
        --seq-length 1024 \
        --max-position-embeddings 1024 \
        --train-iters $MAX_STEPS \
@@ -57,5 +59,6 @@ python -m torch.distributed.launch $DISTRIBUTED_ARGS \
        --eval-iters 10 \
        --tensor-model-parallel-size $TP_SIZE \
        --pipeline-model-parallel-size $PP_SIZE \
+       ${VP_SIZE:+--num-layers-per-virtual-pipeline-stage "$VP_SIZE"} \
        --no-gradient-accumulation-fusion \
        --fp16

tests/functional_tests/test_scripts/gpt3/sbatch_gpt3_distributed_test.sh

@@ -13,4 +13,4 @@ TENSORBOARD_DIR=/workspace/logs
 srun --output $BASE_DIR/results/slurm-%j.out --error $BASE_DIR/results/slurm-%j.out --container-image gitlab-master.nvidia.com/dl/dgx/pytorch:21.12-py3-devel --container-mounts $BASE_DIR/logs:/workspace/logs,$BASE_DIR/checkpoints:/workspace/checkpoints,$BUILD_DIR:/workspace/megatron-lm,$DATA_DIR:/workspace/data --no-container-mount-home bash -c "
   ls 
   cd /workspace/megatron-lm
-  ./tests/functional_tests/test_scripts/gpt3/pretrain_gpt3_distributed_test.sh $DATA_PATH $CHECKPOINT_PATH $TENSORBOARD_DIR $TP_SIZE $PP_SIZE $NUM_NODES $MAX_STEPS"
+  ./tests/functional_tests/test_scripts/gpt3/pretrain_gpt3_distributed_test.sh $DATA_PATH $CHECKPOINT_PATH $TENSORBOARD_DIR $TP_SIZE $PP_SIZE $NUM_NODES $MAX_STEPS $VP_SIZE $MBS $GBS"
\ No newline at end of file