Conversation
This refactoring centralizes model-specific configurations within the provider_bridge method of each model bridge. Changes: - Add MoE-related field mappings to base class CONFIG_MAPPING: - num_experts -> num_moe_experts - num_experts_per_tok -> moe_router_topk - moe_intermediate_size -> moe_ffn_hidden_size - Refactor LlamaBridge: - Use MEGATRON_DEFAULTS and HF_DEFAULTS class attributes - Override provider_bridge only for RoPE scaling (Llama 3.1/3.2) - Refactor Qwen2Bridge: - Use MEGATRON_DEFAULTS (add_qkv_bias=True) and HF_DEFAULTS - No provider_bridge override needed - Refactor Qwen3Bridge: - Use MEGATRON_DEFAULTS (qk_layernorm=True) and HF_DEFAULTS - No provider_bridge override needed - Refactor Qwen3MoEBridge: - Use MEGATRON_DEFAULTS with MoE settings and HF_DEFAULTS - No provider_bridge override needed - Update tests to expect GPTModelProvider instead of model-specific providers - Add verification scripts for both Llama and Qwen bridges Verified on remote server: - Qwen/Qwen2-0.5B: PASS - Qwen/Qwen2-7B: PASS - Qwen/Qwen3-0.6B: PASS - Qwen/Qwen3-1.7B: PASS - Qwen/Qwen3-30B-A3B: PASS
…dels - Add MLAModelProvider as unified base for Multi-Latent Attention models - Refactor DeepSeek V2/V3 bridges to use MLAModelProvider - Refactor Kimi K2 bridge to use MLAModelProvider - Move model-specific defaults from providers to MEGATRON_DEFAULTS in bridges - Add model_type parameter to @register_bridge decorator for auto HF config - Simplify provider files to deprecated backward-compatible aliases Verified: DeepSeek-V2-Lite, DeepSeek-V2, DeepSeek-V3, Moonlight-16B, Kimi-K2
- Register GemmaModelProvider, Gemma2ModelProvider, Gemma3ModelProvider via decorator
- Add MEGATRON_DEFAULTS to Gemma/Gemma2 bridges for explicit config defaults
- Add gelu_pytorch_tanh -> fast_gelu to ACTIVATION_MAPPING in model_bridge.py
- Add verification script for Gemma provider refactoring
Verified: gemma-2b, gemma-7b, gemma-2-2b, gemma-2-9b, gemma-2-27b,
gemma-3-4b-it, gemma-3-12b-it, gemma-3-27b-it
Signed-off-by: yaoyu-33 <yaoyu.094@gmail.com>
Signed-off-by: yaoyu-33 <yaoyu.094@gmail.com>
Signed-off-by: yaoyu-33 <yaoyu.094@gmail.com>
Signed-off-by: yaoyu-33 <yaoyu.094@gmail.com>
Signed-off-by: yaoyu-33 <yaoyu.094@gmail.com>
Signed-off-by: yaoyu-33 <yaoyu.094@gmail.com>
Signed-off-by: yaoyu-33 <yaoyu.094@gmail.com>
Signed-off-by: yaoyu-33 <yaoyu.094@gmail.com>
Signed-off-by: yaoyu-33 <yaoyu.094@gmail.com>
Signed-off-by: yaoyu-33 <yaoyu.094@gmail.com>
OLMoE HF config doesn't have head_dim attribute, so kv_channels was left as None. This fix calculates it as hidden_size // num_attention_heads (2048 // 16 = 128 for OLMoE-1B-7B). This follows the pattern used by MistralBridge and NemotronHBridge.
Signed-off-by: yaoyu-33 <yaoyu.094@gmail.com>
Signed-off-by: yaoyu-33 <yaoyu.094@gmail.com>
Signed-off-by: yaoyu-33 <yaoyu.094@gmail.com>
Signed-off-by: yaoyu-33 <yaoyu.094@gmail.com>
Signed-off-by: yaoyu-33 <yaoyu.094@gmail.com>
Signed-off-by: yaoyu-33 <yaoyu.094@gmail.com>
cuichenx
approved these changes
Feb 12, 2026
This was referenced Feb 14, 2026
5 tasks
This was referenced Feb 24, 2026
Closed
Open
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
2 participants
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.
What does this PR do ?
Add a one line overview of what this PR aims to accomplish.
Changelog
GitHub Actions CI
See the CI sectionin the Contributing doc for how to trigger the CI. A Nvidia developer will need to approve and trigger the CI for external contributors.
Before your PR is "Ready for review"
Pre checks:
If you haven't finished some of the above items you can still open "Draft" PR.
Additional Information
Summary by CodeRabbit
Bug Fixes
generation_configpropagation from model provider bridges to prevent configuration conflicts.New Features
squared_reluactivation function support for enhanced model conversion.Refactor
Tests