Draft
Conversation
- Add activation checkpoint offload module - Add memory defragmentation utilities - Add deep memory profiler script - Add various Kimi 1T training configs (EP64, EP96, EP128, CP2, etc.) - Add Qwen3 activation offload test configs - Add slurm launch scripts - Update DeepSeek V3 model with MoE improvements Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Remove kimi_1t, debug, exp1a training configs, qwen3 test configs, and root-level debug scripts from git tracking.
xrsrke
force-pushed
the
phuc/kimi1t_training
branch
from
9cfe11d to
e04c0f6
Compare
…control - fsdp_reshard_after_forward now accepts integer N for partial resharding to N-GPU groups (e.g., N=8 for intra-node NVLink communication) - Add fsdp_bucket_cap_mb config to control gradient reduction bucket size - Add fsdp_disable_prefetch config to disable forward/backward prefetching - Pass new options through to apply_fsdp() in deepseek_v3 and llama4
- Add nvidia-smi memory reporting for verification against PyTorch stats - Display active memory (actual tensor usage) as primary metric instead of reserved - Log detailed memory breakdown (active/reserved/nvidia-smi) on rank 0 - Enable profile_memory=True in profiler to track allocations per operation
- New AggressiveMemoryManager with 4 modes: minimal, balanced, aggressive, maximum - Clears CUDA cache at strategic points (post-backward, post-optimizer) - Add aggressive_memory_mode and aggressive_memory_verbose config options - Integrate into training loop with post_backward(), post_optimizer(), step_complete() hooks
- Add model presets: qwen3 (2048 dim, 128 experts) and kimi_k2 (7168 dim, 384 experts) - Add init_dist_torchrun() for torchrun environment compatibility - Add CLI arguments: --model, --hidden, --num-experts, --num-topk - Change from group-based to uniform token distribution for routing - Fix MASTER_PORT to be consistent across ranks
- Visualize GPU allocation across DP, PP, TP, CP, EP dimensions - Show mesh structure, submeshes, and coordinate mappings - Visualize expert parallel and context parallel group allocation - Display FSDP sharding details for expert vs non-expert parameters
- Add return_outputs parameter for PP compatibility - Accept **kwargs to handle additional PP arguments
- kimi_k2_12n_ep96_cp16_32k_ctx_lbs11.toml: 12-node baseline config - EP=96, CP=16, DP=1, LBS=11, 32k context - Expected: 402 TPS, 67.55 GiB (85.2%), 17.72% MFU - kimi_k2_36n_ep96_cp16_32k_ctx_hsdp_replicate3_shard6_lbs10.toml: 36-node HSDP config - EP=96, CP=16, dp_replicate=3, dp_shard=6, LBS=10, 32k context - Expected: 378 TPS, 69.45 GiB (87.6%), 16.64% MFU Both configs include aggressive memory management (mode=maximum). Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
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.
No description provided.