[WIP]Enable DP-to-EP for MoE inference#3171
Closed
wwwjn wants to merge 9 commits intogh/wwwjn/16/basefrom
Closed
Conversation
Map vLLM's data_parallel_size to dp_shard in TorchTitan's mesh math, enabling EP to span both DP and TP ranks (ep = dp * tp). For inference, the skip_dp path returns the model directly — no FSDP wrapping needed since there's no backward pass. Changes: - parallel_dims: dp_replicate mesh always exists - vllm_wrapper: ep_size = dp_size * tp_size, dp_shard = dp_size - vllm_wrapper: weight loading uses [Replicate()] * mesh.ndim - qwen3/parallelize: skip_dp returns model (no fully_shard) - llama4/parallelize: clarify shard_placement_fn comments - actors/utils: full_tensor() instead of to_local() for TP-sharded logits - actors/trainer: strict=False (expert_bias not in HF), ep_enabled clip - grpo: tyro.conf.Suppress on model_spec to bypass CLI parsing of Placement - experiments/rl: TORCHTITAN_SKIP_INITIAL_HF_LOAD env var - config_registry: env/validation_env in MoE configs (fix tyro CLI) - qwen3/__init__: debugmodel_moe vocab_size 2048 → 151936 (match tokenizer) - scripts/rl/create_debug_moe_ckpt.py: helper to generate debug checkpoint Verified end-to-end RL with rl_grpo_qwen3_moe_debug_ep on debug model. [ghstack-poisoned]
meta-cla
Bot
added
the
CLA Signed
This was referenced Apr 30, 2026
wwwjn
changed the title
Map vLLM's data_parallel_size to dp_shard in TorchTitan's mesh math, enabling EP to span both DP and TP ranks (ep = dp * tp). For inference, the skip_dp path returns the model directly — no FSDP wrapping needed since there's no backward pass. Changes: - parallel_dims: dp_replicate mesh always exists - vllm_wrapper: ep_size = dp_size * tp_size, dp_shard = dp_size - vllm_wrapper: weight loading uses [Replicate()] * mesh.ndim - qwen3/parallelize: skip_dp returns model (no fully_shard) - llama4/parallelize: clarify shard_placement_fn comments - actors/utils: full_tensor() instead of to_local() for TP-sharded logits - actors/trainer: strict=False (expert_bias not in HF), ep_enabled clip - grpo: tyro.conf.Suppress on model_spec to bypass CLI parsing of Placement - experiments/rl: TORCHTITAN_SKIP_INITIAL_HF_LOAD env var - config_registry: env/validation_env in MoE configs (fix tyro CLI) - qwen3/__init__: debugmodel_moe vocab_size 2048 → 151936 (match tokenizer) - scripts/rl/create_debug_moe_ckpt.py: helper to generate debug checkpoint Verified end-to-end RL with rl_grpo_qwen3_moe_debug_ep on debug model. [ghstack-poisoned]
Map vLLM's data_parallel_size to dp_shard in TorchTitan's mesh math, enabling EP to span both DP and TP ranks (ep = dp * tp). For inference, the skip_dp path returns the model directly — no FSDP wrapping needed since there's no backward pass. Changes: - parallel_dims: dp_replicate mesh always exists - vllm_wrapper: ep_size = dp_size * tp_size, dp_shard = dp_size - vllm_wrapper: weight loading uses [Replicate()] * mesh.ndim - qwen3/parallelize: skip_dp returns model (no fully_shard) - llama4/parallelize: clarify shard_placement_fn comments - actors/utils: full_tensor() instead of to_local() for TP-sharded logits - actors/trainer: strict=False (expert_bias not in HF), ep_enabled clip - grpo: tyro.conf.Suppress on model_spec to bypass CLI parsing of Placement - experiments/rl: TORCHTITAN_SKIP_INITIAL_HF_LOAD env var - config_registry: env/validation_env in MoE configs (fix tyro CLI) - qwen3/__init__: debugmodel_moe vocab_size 2048 → 151936 (match tokenizer) - scripts/rl/create_debug_moe_ckpt.py: helper to generate debug checkpoint Verified end-to-end RL with rl_grpo_qwen3_moe_debug_ep on debug model. [ghstack-poisoned]
Map vLLM's data_parallel_size to dp_shard in TorchTitan's mesh math, enabling EP to span both DP and TP ranks (ep = dp * tp). For inference, the skip_dp path returns the model directly — no FSDP wrapping needed since there's no backward pass. Changes: - parallel_dims: dp_replicate mesh always exists - vllm_wrapper: ep_size = dp_size * tp_size, dp_shard = dp_size - vllm_wrapper: weight loading uses [Replicate()] * mesh.ndim - qwen3/parallelize: skip_dp returns model (no fully_shard) - llama4/parallelize: clarify shard_placement_fn comments - actors/utils: full_tensor() instead of to_local() for TP-sharded logits - actors/trainer: strict=False (expert_bias not in HF), ep_enabled clip - grpo: tyro.conf.Suppress on model_spec to bypass CLI parsing of Placement - experiments/rl: TORCHTITAN_SKIP_INITIAL_HF_LOAD env var - config_registry: env/validation_env in MoE configs (fix tyro CLI) - qwen3/__init__: debugmodel_moe vocab_size 2048 → 151936 (match tokenizer) - scripts/rl/create_debug_moe_ckpt.py: helper to generate debug checkpoint Verified end-to-end RL with rl_grpo_qwen3_moe_debug_ep on debug model. [ghstack-poisoned]
Map vLLM's data_parallel_size to dp_shard in TorchTitan's mesh math, enabling EP to span both DP and TP ranks (ep = dp * tp). For inference, the skip_dp path returns the model directly — no FSDP wrapping needed since there's no backward pass. Changes: - parallel_dims: dp_replicate mesh always exists - vllm_wrapper: ep_size = dp_size * tp_size, dp_shard = dp_size - vllm_wrapper: weight loading uses [Replicate()] * mesh.ndim - qwen3/parallelize: skip_dp returns model (no fully_shard) - llama4/parallelize: clarify shard_placement_fn comments - actors/utils: full_tensor() instead of to_local() for TP-sharded logits - actors/trainer: strict=False (expert_bias not in HF), ep_enabled clip - grpo: tyro.conf.Suppress on model_spec to bypass CLI parsing of Placement - experiments/rl: TORCHTITAN_SKIP_INITIAL_HF_LOAD env var - config_registry: env/validation_env in MoE configs (fix tyro CLI) - qwen3/__init__: debugmodel_moe vocab_size 2048 → 151936 (match tokenizer) - scripts/rl/create_debug_moe_ckpt.py: helper to generate debug checkpoint Verified end-to-end RL with rl_grpo_qwen3_moe_debug_ep on debug model. [ghstack-poisoned]
Map vLLM's data_parallel_size to dp_shard in TorchTitan's mesh math, enabling EP to span both DP and TP ranks (ep = dp * tp). For inference, the skip_dp path returns the model directly — no FSDP wrapping needed since there's no backward pass. Changes: - parallel_dims: dp_replicate mesh always exists - vllm_wrapper: ep_size = dp_size * tp_size, dp_shard = dp_size - vllm_wrapper: weight loading uses [Replicate()] * mesh.ndim - qwen3/parallelize: skip_dp returns model (no fully_shard) - llama4/parallelize: clarify shard_placement_fn comments - actors/utils: full_tensor() instead of to_local() for TP-sharded logits - actors/trainer: strict=False (expert_bias not in HF), ep_enabled clip - grpo: tyro.conf.Suppress on model_spec to bypass CLI parsing of Placement - experiments/rl: TORCHTITAN_SKIP_INITIAL_HF_LOAD env var - config_registry: env/validation_env in MoE configs (fix tyro CLI) - qwen3/__init__: debugmodel_moe vocab_size 2048 → 151936 (match tokenizer) - scripts/rl/create_debug_moe_ckpt.py: helper to generate debug checkpoint Verified end-to-end RL with rl_grpo_qwen3_moe_debug_ep on debug model. [ghstack-poisoned]
Map vLLM's data_parallel_size to dp_shard in TorchTitan's mesh math, enabling EP to span both DP and TP ranks (ep = dp * tp). For inference, the skip_dp path returns the model directly — no FSDP wrapping needed since there's no backward pass. Changes: - parallel_dims: dp_replicate mesh always exists - vllm_wrapper: ep_size = dp_size * tp_size, dp_shard = dp_size - vllm_wrapper: weight loading uses [Replicate()] * mesh.ndim - qwen3/parallelize: skip_dp returns model (no fully_shard) - llama4/parallelize: clarify shard_placement_fn comments - actors/utils: full_tensor() instead of to_local() for TP-sharded logits - actors/trainer: strict=False (expert_bias not in HF), ep_enabled clip - grpo: tyro.conf.Suppress on model_spec to bypass CLI parsing of Placement - experiments/rl: TORCHTITAN_SKIP_INITIAL_HF_LOAD env var - config_registry: env/validation_env in MoE configs (fix tyro CLI) - qwen3/__init__: debugmodel_moe vocab_size 2048 → 151936 (match tokenizer) - scripts/rl/create_debug_moe_ckpt.py: helper to generate debug checkpoint Verified end-to-end RL with rl_grpo_qwen3_moe_debug_ep on debug model. [ghstack-poisoned]
Map vLLM's data_parallel_size to dp_shard in TorchTitan's mesh math, enabling EP to span both DP and TP ranks (ep = dp * tp). For inference, the skip_dp path returns the model directly — no FSDP wrapping needed since there's no backward pass. Changes: - parallel_dims: dp_replicate mesh always exists - vllm_wrapper: ep_size = dp_size * tp_size, dp_shard = dp_size - vllm_wrapper: weight loading uses [Replicate()] * mesh.ndim - qwen3/parallelize: skip_dp returns model (no fully_shard) - llama4/parallelize: clarify shard_placement_fn comments - actors/utils: full_tensor() instead of to_local() for TP-sharded logits - actors/trainer: strict=False (expert_bias not in HF), ep_enabled clip - grpo: tyro.conf.Suppress on model_spec to bypass CLI parsing of Placement - experiments/rl: TORCHTITAN_SKIP_INITIAL_HF_LOAD env var - config_registry: env/validation_env in MoE configs (fix tyro CLI) - qwen3/__init__: debugmodel_moe vocab_size 2048 → 151936 (match tokenizer) - scripts/rl/create_debug_moe_ckpt.py: helper to generate debug checkpoint Verified end-to-end RL with rl_grpo_qwen3_moe_debug_ep on debug model. [ghstack-poisoned]
wwwjn
added a commit
that referenced
this pull request
May 5, 2026
…cher (#3193) Stack from [ghstack](https://github.com/ezyang/ghstack) (oldest at bottom): * #3172 * #3171 * #3142 * __->__ #3193 Today the dispatcher's _split_along_sp() raises when num_tokens (bs * slen) is not divisible by sp_size (= TP degree). Real workloads with varlen prompts can land on non-divisible totals and crash the MoE forward. Pad inside dispatch(): round num_tokens up to the next multiple of sp_size, padding x and top_scores with zeros and selected_experts_indices with 0 (so pad rows route deterministically to expert 0 with zero score). combine() reads metadata.original_num_tokens to size the scatter_add buffer at the padded length and slices the pad rows off before returning. When sp_size == 1 or input is already divisible, behavior is bitwise identical to today. Pad tokens are numerically inert: - Zero scores -> contribution to scatter_add is exactly zero either before or after expert compute (independent of score_before_experts). - Pad indices fall in [original, padded), which is sliced off after scatter_add, so they never appear in the returned output. Trainer/generator can pad by different amounts depending on their batch shapes; the unpadded portions remain bitwise identical. - TorchAOTokenDispatcher inherits dispatch/combine and gets the fix for free. - DeepEPTokenDispatcher uses a separate metadata type and is unaffected.
Map vLLM's data_parallel_size to dp_shard in TorchTitan's mesh math, enabling EP to span both DP and TP ranks (ep = dp * tp). For inference, the skip_dp path returns the model directly — no FSDP wrapping needed since there's no backward pass. Changes: - parallel_dims: dp_replicate mesh always exists - vllm_wrapper: ep_size = dp_size * tp_size, dp_shard = dp_size - vllm_wrapper: weight loading uses [Replicate()] * mesh.ndim - qwen3/parallelize: skip_dp returns model (no fully_shard) - llama4/parallelize: clarify shard_placement_fn comments - actors/utils: full_tensor() instead of to_local() for TP-sharded logits - actors/trainer: strict=False (expert_bias not in HF), ep_enabled clip - grpo: tyro.conf.Suppress on model_spec to bypass CLI parsing of Placement - experiments/rl: TORCHTITAN_SKIP_INITIAL_HF_LOAD env var - config_registry: env/validation_env in MoE configs (fix tyro CLI) - qwen3/__init__: debugmodel_moe vocab_size 2048 → 151936 (match tokenizer) - scripts/rl/create_debug_moe_ckpt.py: helper to generate debug checkpoint Verified end-to-end RL with rl_grpo_qwen3_moe_debug_ep on debug model. [ghstack-poisoned]
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Stack from ghstack (oldest at bottom):
Map vLLM's data_parallel_size to dp_shard in TorchTitan's mesh math,
enabling EP to span both DP and TP ranks (ep = dp * tp). For inference,
the skip_dp path returns the model directly — no FSDP wrapping needed
since there's no backward pass.
Changes:
Verified end-to-end RL with rl_grpo_qwen3_moe_debug_ep on debug model.