Functions — github.com/VizuaraAILabs/truly-open-gpt-oss

CodeHub MCPcopy Create free account

hub / github.com/VizuaraAILabs/truly-open-gpt-oss / functions

Functions69 in github.com/VizuaraAILabs/truly-open-gpt-oss

⨍ Functions 69 ◇ Types & classes 9

↓ 19 callersFunctionrank0_print

(*args, **kwargs)

↓ 13 callersFunctionis_dist

↓ 9 callersFunctionis_dist

export_to_safetensors.py:50

↓ 8 callersMethodreset_parameters

↓ 7 callersFunctionis_dist

↓ 5 callersMethod__init__

(self, cfg: ModelConfig)

↓ 3 callersMethodget_batch

↓ 3 callersFunctionrank0_print

(*args, **kwargs)

↓ 3 callersFunctionsharded_ckpt_path

(prefix: str, rank: int)

↓ 3 callersFunctionsharded_path

(in_dir, prefix, rank)

export_to_safetensors.py:57

↓ 2 callersMethod_kv_expand

(self, kv: torch.Tensor)

↓ 2 callersFunction_massage_cfg_for_model

Turn nested dicts into expected types (e.g., rope_scaling dict -> RopeScalingConfig).

export_to_safetensors.py:79

↓ 2 callersMethodapply

(self, x: torch.Tensor, positions: torch.Tensor)

↓ 2 callersFunctionencode_corpus

(texts: List[str], enc)

↓ 2 callersFunctiongpt_oss_20b_config

↓ 2 callersFunctionrank0_print

export_to_safetensors.py:53

↓ 2 callersFunctionsave_ckpt_sharded

(prefix: str, iter_n: int, best_v: float)

↓ 2 callersFunctionsharded_path

(out_dir, prefix, rank)

↓ 2 callersFunctionwrite_memmap

(path: str, tokens: np.ndarray)

↓ 1 callersMethod_build_local_mask

(T: int, device, win: int)

↓ 1 callersFunction_massage_cfg_for_model

Normalize nested fields so model.py (RotaryEmbedding etc.) gets attribute-like objects.

↓ 1 callersMethod_update_cache

(self, seqlen: int, device, dtype)

↓ 1 callersFunctionapply_repetition_penalty

(logits, generated_ids, penalty)

↓ 1 callersMethodbuild_causal_mask

(T: int, device, dtype=torch.bool)

↓ 1 callersFunctionbuild_config

↓ 1 callersFunctionbuild_model_from_cfg

(cfg_dict: dict, device: str, args)

export_to_safetensors.py:94

↓ 1 callersFunctionbuild_model_from_config

(cfg: dict, device: str, args)

↓ 1 callersFunctionevaluate

↓ 1 callersFunctiongenerate_collective

All ranks participate in forward passes (FSDP requirement). Rank 0 samples next token id and broadcasts to all ranks so sequences stay in syn

↓ 1 callersFunctionget_args

↓ 1 callersFunctionget_args

↓ 1 callersFunctionget_args

export_to_safetensors.py:37

↓ 1 callersFunctionget_lr

↓ 1 callersFunctionget_o200k_harmony_tokenizer

Try to get 'o200k_harmony'. If not present, fall back to 'o200k_base' and warn, but write the actual tokenizer name into meta.json so the

↓ 1 callersFunctioninit_dist

↓ 1 callersFunctionload_ckpt_full

(single_path: str)

↓ 1 callersFunctionload_ckpt_sharded

↓ 1 callersFunctionload_tokenizer

(meta_path: str)

↓ 1 callersFunctionmain

↓ 1 callersFunctionmain

↓ 1 callersFunctionmain

↓ 1 callersFunctionmain

export_to_safetensors.py:170

↓ 1 callersFunctionparse_size

export_to_safetensors.py:61

↓ 1 callersMethodreset_parameters

↓ 1 callersMethodreset_parameters

↓ 1 callersMethodreset_parameters

↓ 1 callersFunctionsample_text_collective

All ranks enter (so FSDP collectives don't deadlock). Only rank 0 returns the decoded text; others return "".

↓ 1 callersFunctionshard_state_dict_for_hf

Split a full weights dict into shards <= max_shard_bytes. Returns: shards: { filename -> {param_name: tensor, ...} } index: {"me

export_to_safetensors.py:127

↓ 1 callersFunctionswiglu

(x: torch.Tensor, clip: Optional[float] = None)

↓ 1 callersFunctiontop_k_filter

↓ 1 callersFunctiontop_p_filter

(logits, top_p)

(self, data_dir: str, split: str, block_size: int, batch_size: int, device: str, seed: int)

(self, dim: int, eps: float = 1e-5)

(self, head_dim: int, rope_theta: float, scale_cfg: RopeScalingConfig)

(self, cfg: ModelConfig)

(self, cfg: ModelConfig)

(self, cfg: ModelConfig)

Method__post_init__

Function_fsdp_param_init_fn

(m: torch.nn.Module)

Function_param_init_fn

(m: torch.nn.Module)

Function_param_init_fn

(m: torch.nn.Module)

export_to_safetensors.py:111

(self, x: torch.Tensor)

( self, x: torch.Tensor, # (B,T,H) positions: torch.Tensor, # (B,T

(self, x: torch.Tensor)

( self, x: torch.Tensor, positions: torch.Tensor, causal_mask: torch.Tensor,

( self, input_ids: torch.Tensor, # (B,T) labels: Optional[torch.Tenso

Methodgroup_size

Methodreset_parameters

Methodreset_parameters