Add GLM-4.7-Flash (30B MoE) training example

[tyler-griggs]

Summary

Adds a complete GRPO training example for GLM-4.7-Flash (zai-org/GLM-4.7-Flash), a DeepSeek-V3 architecture clone with MLA + MoE (64 routed experts, 4 active per token, ~3B active parameters).

Files

• skyrl/train/config/examples/glm4_7_30b_moe_grpo.yaml — Hydra config
• skyrl-train/examples/megatron/run_megatron_grpo_glm4_7_30b.sh — Launch script

Configuration

• Parallelism: TP=2, EP=8 on 2 nodes x 8 H100 GPUs (16 GPUs total)
• MoE flags: sigmoid routing, expert bias, grouped GEMM
• Attention: flash_attn: false (MLA not compatible with TE flash attention)
• Algorithm: GRPO with KL loss

Dependencies

Requires the following PRs to be merged first:

• Add configurable MoE runtime flags to MegatronConfig #1213 (MoE config fields)
• Register GLM-4.7-Flash bridge and bump megatron-bridge #1214 (GLM-4.7-Flash bridge registration)

---

[gemini-code-assist]

Code Review

This pull request adds a new training example for GLM-4.7-Flash. The changes look good overall, but there are a couple of issues to address. First, the launch script duplicates a lot of configuration from the YAML file, which is a maintainability concern. I've suggested refactoring the script to load the YAML file directly. Second, the generator configuration in the YAML file is inconsistent with the documented 8-GPU setup, as it's configured for 16 GPUs. I've suggested correcting these values.

[gemini-code-assist]

The uv run command duplicates a large amount of configuration that is also present in the glm4_7_30b_moe_grpo.yaml file. This makes maintenance difficult, as changes need to be synchronized in two places. The script should be simplified to load the YAML configuration file using Hydra's command-line arguments and only override parameters that are specific to the execution environment.

This would make the script much shorter and more maintainable. For example, you could use --config-name to specify the example config. You might need to also use --config-dir or --config-path if the example config is not in Hydra's default search path.

A refactored command could look like this:

# Assuming the config path is correctly set up for Hydra to find the example config.
CONFIG_NAME="examples/glm4_7_30b_moe_grpo"

uv run --isolated --extra mcore -m skyrl_train.entrypoints.main_base \
  --config-name $CONFIG_NAME \
  data.train_data="['$DATA_DIR/train.parquet']" \
  data.val_data="['$DATA_DIR/validation.parquet']" \
  trainer.placement.policy_num_nodes=$NUM_NODES \
  trainer.placement.ref_num_nodes=$NUM_NODES \
  trainer.placement.policy_num_gpus_per_node=$NUM_GPUS \
  trainer.placement.ref_num_gpus_per_node=$NUM_GPUS \
  trainer.policy.megatron_config.tensor_model_parallel_size=$MEGATRON_TP \
  trainer.policy.megatron_config.expert_model_parallel_size=$MEGATRON_EP \
  generator.num_inference_engines=$NUM_INFERENCE_ENGINES \
  generator.inference_engine_tensor_parallel_size=$INFERENCE_ENGINE_TP \
  trainer.logger="$LOGGER" \
  trainer.run_name="glm4_7_30b_a3b_grpo_megatron_tp${MEGATRON_TP}_pp${MEGATRON_PP}_cp${MEGATRON_CP}_ep${MEGATRON_EP}_etp${MEGATRON_ETP}" \
  trainer.ckpt_path="$HOME/ckpts/glm4_7_30b_a3b_grpo_megatron" \
  $@

Most of the parameters from line 46 to 106 could be removed from the script as they would be loaded from the YAML file. I've kept a few as an example of what might be useful to keep as overrides.

[gemini-code-assist]

The generator configuration seems to be incorrect for the intended hardware setup. The comments state this configuration is for an 8-GPU setup, but the generator is configured with num_inference_engines: 2 and inference_engine_tensor_parallel_size: 8, which would require 16 GPUs (2 * 8). This contradicts the comment and the setup in the corresponding launch script run_megatron_grpo_glm4_7_30b.sh, which uses more reasonable values for an 8-GPU node (NUM_INFERENCE_ENGINES=1, INFERENCE_ENGINE_TP=4). Please update these values to be consistent with an 8-GPU setup.

  num_inference_engines: 1
  inference_engine_tensor_parallel_size: 4

[devin-ai-integration]

Devin Review found 2 potential issues.

View 3 additional findings in Devin Review.

[devin-ai-integration]

🔴 YAML inference engine GPU count mismatches policy GPU count, causing assertion failure

The YAML config sets colocate_all: true with policy_num_nodes: 1 and policy_num_gpus_per_node: 8 (8 policy GPUs), but num_inference_engines: 2 and inference_engine_tensor_parallel_size: 8 (16 rollout GPUs). This will trigger the assertion at skyrl/train/utils/utils.py:363 which requires num_policy_gpus == num_rollout_gpus when colocate_all is true.

Root Cause

The config appears to have been copied from a 2-node setup (like the Qwen3-30B or Moonlight examples which use 2 nodes × 8 GPUs = 16 GPUs with num_inference_engines=2, inference_engine_tensor_parallel_size=8) without adjusting the inference engine settings for the 1-node/8-GPU target.

The validation at skyrl/train/utils/utils.py:354-366 computes:

• num_policy_gpus = 1 * 8 = 8
• num_rollout_gpus = 2 * 8 * 1 * 1 = 16

Since 8 != 16, the assertion fails and training cannot start.

Impact: The YAML config is unusable as-is — it will crash immediately at config validation.

Suggested change

	num_inference_engines: 2
	inference_engine_tensor_parallel_size: 8
	num_inference_engines: 1
	inference_engine_tensor_parallel_size: 8

---

Was this helpful? React with 👍 or 👎 to provide feedback.

[devin-ai-integration]

🔴 Shell script inference engine GPU count mismatches policy GPU count, causing assertion failure

The shell script sets colocate_all=true with NUM_NODES=1 and NUM_GPUS=8 (8 policy GPUs), but NUM_INFERENCE_ENGINES=1 and INFERENCE_ENGINE_TP=4 (4 rollout GPUs). This will trigger the assertion at skyrl/train/utils/utils.py:363 which requires num_policy_gpus == num_rollout_gpus when colocate_all is true.

Root Cause

The validation at skyrl/train/utils/utils.py:354-366 computes:

• num_policy_gpus = 1 * 8 = 8
• num_rollout_gpus = 1 * 4 * 1 * 1 = 4

Since 8 != 4, the assertion fails. Comparing with working examples like run_megatron_moonlight.sh and run_megatron_qwen3-30b-a3b.sh which both use NUM_INFERENCE_ENGINES=2, INFERENCE_ENGINE_TP=8 for 16 GPUs (matching their NUM_NODES=2 * NUM_GPUS=8 = 16), the correct values for 1 node × 8 GPUs should be either NUM_INFERENCE_ENGINES=1, INFERENCE_ENGINE_TP=8 or NUM_INFERENCE_ENGINES=2, INFERENCE_ENGINE_TP=4.

Impact: The shell script will crash immediately at config validation and cannot be used to launch training.

Suggested change

	NUM_INFERENCE_ENGINES=1
	INFERENCE_ENGINE_TP=4
	NUM_INFERENCE_ENGINES=1
	INFERENCE_ENGINE_TP=8

---

Was this helpful? React with 👍 or 👎 to provide feedback.