Add support for Accelerate CP, ND examples, and fix for parallel config w fsdp (#3019)

* fix for parallelism config from trainer

* fix handling of parallelism_config w accelerate

* add todo for removal

* update to latest axolotl-contribs-mit for optimizer fix too

* synchronize training after checkpoint save

* dir spelling

* use latest accelerate main

* fix to not use partial state parallelism_config

* more fixeS

* use most recent accelerate fix

* fix cpu_ram_efficient_loading to meta devices from rank 0 to prevent CPU RAM oom

* improve handling of broadcasting fsdp2 state dict

* support for openai chat template with thinking key as the reasoning trace

* address PR feedback

* refactor to remove dependency on PartialState for parallelism config

* bump accelerate, gptoss fixes

* limit meta fixes to fsdp2 for now

* fixes for gpt oss

* fixup examples, don't use cpu-ram-efficient-loading for now

* remove problematic barrier

* patch parallelism config

* reorder comparison

* device mesh fixes

* make pure CP work

* lint

examples/distributed-parallel/README.md

@@ -0,0 +1,8 @@
+# Distributed Parallel
+
+See the accompanying blog post: [Accelerate ND-Parallel: A guide to Efficient Multi-GPU Training](https://huggingface.co/blog/accelerate-nd-parallel)
+
+The examples provided are suitable for single node (8xGPU) SFT.
+
+- Qwen 3 8B w/ FSDP + TP + CP: [YAML](./qwen3-8b-fsdp-tp-cp.yaml)
+- Llama 3.1 8B w/ HSDP + TP: [YAML](./llama-3_1-8b-hdsp-tp.yaml)

examples/distributed-parallel/llama-3_1-8b-hdsp-tp.yaml

@@ -0,0 +1,47 @@
+base_model: meta-llama/Llama-3.1-8B
+
+plugins:
+  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin
+
+dp_shard_size: 2
+dp_replicate_size: 2
+tensor_parallel_size: 2
+# context_parallel_size: 2
+
+dataset_prepared_path: last_run_prepared
+
+special_tokens:
+  pad_token: <|end_of_text|>
+
+fsdp_version: 2
+fsdp_config:
+  offload_params: false
+  state_dict_type: FULL_STATE_DICT
+  auto_wrap_policy: TRANSFORMER_BASED_WRAP
+  transformer_layer_cls_to_wrap: LlamaDecoderLayer
+  reshard_after_forward: true
+
+datasets:
+  - path: tatsu-lab/alpaca
+    type: alpaca
+
+output_dir: ./outputs/ndp-out/
+
+sequence_len: 2048
+sample_packing: true
+flash_attention: true
+
+gradient_accumulation_steps: 1
+micro_batch_size: 1
+num_epochs: 2
+optimizer: adamw_torch_fused
+lr_scheduler: constant_with_warmup
+learning_rate: 2e-6
+
+bf16: true
+tf32: true
+
+logging_steps: 1
+saves_per_epoch: 1
+
+warmup_ratio: 0.1

examples/distributed-parallel/qwen3-8b-fsdp-tp-cp.yaml

@@ -0,0 +1,46 @@
+base_model: Qwen/Qwen3-8B
+
+plugins:
+  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin
+
+dp_shard_size: 2
+# dp_replicate_size: 1
+context_parallel_size: 2
+tensor_parallel_size: 2
+
+dataset_prepared_path: last_run_prepared
+
+fsdp_version: 2
+fsdp_config:
+  offload_params: false
+  state_dict_type: FULL_STATE_DICT
+  auto_wrap_policy: TRANSFORMER_BASED_WRAP
+  transformer_layer_cls_to_wrap: Qwen3DecoderLayer
+  reshard_after_forward: true
+
+datasets:
+  - path: tatsu-lab/alpaca
+    type: alpaca
+
+output_dir: ./outputs/ndp-out/
+
+sequence_len: 8192
+sample_packing: true
+flash_attention: true
+
+gradient_accumulation_steps: 1
+micro_batch_size: 1  # must be 1 when using context parallel
+num_epochs: 2
+optimizer: adamw_torch_fused
+lr_scheduler: constant_with_warmup
+learning_rate: 2e-6
+
+bf16: true
+tf32: true
+
+logging_steps: 1
+saves_per_epoch: 1
+
+warmup_ratio: 0.1
+
+special_tokens: