Checkpoint File → from_pretrained() → convert_and_load_state_dict_in_model()
                                              ↓
                         ┌───────────────────────────────────────────────────────────┐
                         │  For each weight in checkpoint:                           │
                         │  1. Match renamed/processed source key to model parameter │
                         │  2. Shard the weight and send to device (async)           │
                         │  3. Collect tensors with the same source_pattern together │
                         │     (e.g. MoE experts, gate_up_proj)                      │
                         │  4. Apply dequantization/deserialization (if pre-quant)   │
                         │  5. Apply conversion (if defined)                         │
                         │  6. Apply quantization (if enabled and step 4 not used)   │
                         │  7. Set parameter on model                                │
                         └───────────────────────────────────────────────────────────┘

Checkpoint:                             Model:
model.layers.0.self_attn.q_proj.weight  →  model.layers.0.self_attn.q_proj.weight
model.layers.0.self_attn.k_proj.weight  →  model.layers.0.self_attn.k_proj.weight
model.layers.0.mlp.gate_proj.weight     →  model.layers.0.mlp.gate_proj.weight
model.layers.0.mlp.up_proj.weight       →  model.layers.0.mlp.up_proj.weight
model.layers.0.mlp.down_proj.weight     →  model.layers.0.mlp.down_proj.weight

Checkpoint:                          Model:
LayerNorm.gamma              →       LayerNorm.weight
LayerNorm.beta               →       LayerNorm.bias

Checkpoint:                              Model:
experts.0.w1.weight  ─┐
experts.1.w1.weight   │ MergeModulelist
...                   ├───────────────→  experts.gate_up_proj (8, hidden, 2*intermediate)
experts.0.w3.weight   │ + Concatenate
experts.1.w3.weight  ─┘

Composite model:                              Per-submodel mappings (auto-scoped):
LlavaForConditionalGeneration
├── vision_model: SiglipVisionModel    →     SiglipVisionModel mapping (scope="vision_model")
└── language_model: LlamaForCausalLM   →     LlamaForCausalLM mapping  (scope="language_model")

# Legacy checkpoint compatibility
WeightRenaming("LayerNorm.gamma", "LayerNorm.weight")

# Module path changes
WeightRenaming(".block_sparse_moe.", ".mlp.")

# "model.layers.bad_prefix.weight" → "model.layers.weight"
PrefixChange(prefix_to_remove="bad_prefix", model_prefix="model.layers")
# "layers.0.weight" → "model.layers.0.weight"
PrefixChange(prefix_to_add="model")

WeightConverter(
    source_patterns=[".experts.*.w1.weight", ".experts.*.w3.weight"],
    target_patterns=".experts.gate_up_proj",
    operations=[MergeModulelist(dim=0), Concatenate(dim=1)],
)

weight_mapping = [
    WeightRenaming("^old_prefix", "encoder"),              # rename runs always
    WeightConverter(                                       # converter claims the key
        "attn.qkv_proj.weight",
        ["attn.q_proj.weight", "attn.k_proj.weight", "attn.v_proj.weight"],
        operations=[Chunk(dim=0)],
    ),
]
# Load:  "old_prefix.attn.qkv_proj.weight"
#   → WeightRenaming  → "encoder.attn.qkv_proj.weight"
#   → WeightConverter → "encoder.attn.{q,k,v}_proj.weight"
#
# Save: list reversed, each transform inverted:
#   → rev(WeightConverter) repacks QKV → "encoder.attn.qkv_proj.weight"
#   → rev(WeightRenaming)  fixes prefix → "old_prefix.attn.qkv_proj.weight"

WeightConverter(
    "self_attn.qkv_proj",
    ["self_attn.q_proj", "self_attn.k_proj", "self_attn.v_proj"],
    operations=[Chunk(dim=0)],
)

WeightConverter(
    ["self_attn.q_proj", "self_attn.k_proj", "self_attn.v_proj"],
    "self_attn.qkv_proj",
    operations=[Concatenate(dim=0)],
)

WeightConverter(
    ["block_sparse_moe.experts.*.w1.weight", "block_sparse_moe.experts.*.w3.weight"],
    "mlp.experts.gate_up_proj",
    operations=[MergeModulelist(dim=0), Concatenate(dim=1)],
)

WeightConverter(
    "mlp.experts.down_proj",
    "block_sparse_moe.experts.*.w2.weight",
    operations=[SplitModulelist(dim=0)],
)

WeightConverter(
    ["model.layers.*.self_attn.qkv_proj.weight"],
    [
        "model.layers.*.self_attn.q_proj.weight",
        "model.layers.*.self_attn.k_proj.weight",
        "model.layers.*.self_attn.v_proj.weight",
    ],
    operations=[Chunk(dim=0), PermuteForRope()],
)

WeightConverter(
    source_patterns="mlp.gate.weight",
    target_patterns="mlp.text_moe.gate.weight",
    operations=[Transpose(dim0=0, dim1=1)],
)

WeightConverter(
    source_patterns=[
        ".experts.*.w1.weight",  # gate_proj per expert
        ".experts.*.w3.weight",  # up_proj per expert
    ],
    target_patterns=".experts.gate_up_proj",
    operations=[
        MergeModulelist(dim=0),  # Stack all experts: (n_experts, in, out)
        Concatenate(dim=1),      # Fuse gate+up: (n_experts, in, 2*out)
    ],
)

Input:
  ".experts.*.w1.weight": [tensor_0, tensor_1, ..., tensor_7]  # 8 experts
  ".experts.*.w3.weight": [tensor_0, tensor_1, ..., tensor_7]  # 8 experts

After MergeModulelist(dim=0):
  ".experts.*.w1.weight": (8, 4096, 14336)  # stacked gate
  ".experts.*.w3.weight": (8, 4096, 14336)  # stacked up

After Concatenate(dim=1):
  ".experts.gate_up_proj": (8, 4096, 28672)  # fused gate_up

from transformers.conversion_mapping import register_checkpoint_conversion_mapping

register_checkpoint_conversion_mapping(
    "my_model_type",
    [
        WeightRenaming(".old.", ".new."),
        WeightConverter(
            ".experts.*.w1.weight",
            ".experts.gate_proj",
            operations=[MergeModulelist(dim=0)],
        ),
    ],
)

_MODEL_TO_CONVERSION_PATTERN = {
    # model_type aliases (lookup by config.model_type)
    "minimax": "mixtral",
    "qwen3_moe": "qwen2_moe",
    "mistral3": "llava",
    # class-name aliases (lookup by type(submodule).__name__)
    "PaliGemmaModel": "LlavaModel",
    "MaskFormerDetrDecoder": "DetrModel",
    ...
}

def spawn_tp_materialize(thread_pool, tensor, sharding_method, tensor_idx, device, dtype):
    def _job():
        return sharding_method.shard_tensor(tensor, tensor_idx=tensor_idx, device=device, dtype=dtype)
    return thread_pool.submit(_job)

def get_packed_weights(param, empty_param, device_mesh, rank, dim):
    """
    Interleaves gate and up shards correctly.

    Packed tensor: [G0 G1 G2 G3 | U0 U1 U2 U3]

    With TP=2:
    - Rank 0 gets: [G0 G1 | U0 U1]
    - Rank 1 gets: [G2 G3 | U2 U3]
    """

if matched_tp_pattern := tp_plan_alt.search(renamed_key):
    tp_layer = ALL_PARALLEL_STYLES[model.tp_plan[matched_tp_pattern]]
    mapping.distributed_operation = tp_layer(
        device_mesh=device_mesh,
        rank=device_mesh.get_local_rank(),
        empty_param=empty_param.clone()
    )

def spawn_materialize(thread_pool, tensor, device, dtype) -> Future | Callable:
    def _job():
        return _materialize_copy(tensor, device, dtype)

    if thread_pool is not None:
        return thread_pool.submit(_job)  # Async: returns Future
    else:
        return _job  # Sync: returns Callable (deferred execution)

1. Checkpoint iteration (Phase 1):
   - For each key, walk the transform list once
   - Submit materialization job to ThreadPoolExecutor
   - Job returns Future (async) or Callable (sync)
   - Collect into the matching WeightConverter / WeightRenaming

2. Per-mapping processing (Phase 2, one mapping at a time):
   - materialize_tensors() waits for this mapping's Futures only
   - Apply conversion operations chain (self.operations)
   - Apply quantization operation (if on-the-fly)
   - Set parameters on model
   - Delete realized tensors immediately

3. Cleanup:
   - Thread pool shutdown (with cancel_futures=True for interrupts)

def revert_weight_conversion(model, state_dict):
    """Applies reverse conversions for saving."""
    weight_conversions = getattr(model, "_weight_conversions", None)

    # Reverse all transforms
    reverse_weight_conversion = [
        conversion.reverse_transform() for conversion in weight_conversions
    ]

    # Apply in reverse
    for first_param_name, reversed_converter in conversion_mapping.items():
        realized_value = reversed_converter.convert(first_param_name, model=model)

def process_target_pattern(pattern: str) -> tuple[str, str | None]:
    """
    - Removes `^` and `$` anchors
    - Removes negative lookahead/lookbehind
    - Detects capturing groups, replaces with \1
    """

model.layers.0.block_sparse_moe.experts.0.w1.weight  # gate per expert
model.layers.0.block_sparse_moe.experts.0.w2.weight  # down per expert
model.layers.0.block_sparse_moe.experts.0.w3.weight  # up per expert
...
model.layers.0.block_sparse_moe.experts.7.w1.weight

model.layers.0.mlp.experts.gate_up_proj  # (8, 4096, 28672)
model.layers.0.mlp.experts.down_proj     # (8, 14336, 4096)

"mixtral": [
    WeightRenaming(".block_sparse_moe.", ".mlp."),
    WeightConverter(
        source_patterns=[".experts.*.w1.weight", ".experts.*.w3.weight"],
        target_patterns=".experts.gate_up_proj",
        operations=[MergeModulelist(dim=0), Concatenate(dim=1)],
    ),
    WeightConverter(
        source_patterns=[".experts.*.w2.weight"],
        target_patterns=".experts.down_proj",
        operations=[MergeModulelist(dim=0)],
    ),
],

# model_type: applies to Gemma3ForConditionalGeneration / Gemma3ForSequenceClassification
"gemma3": "llava"           # → adds "model." in front of language_model / vision_tower / ...

# class:    applies to the inner Gemma3Model only
"Gemma3Model": "LlavaModel"  # → minimal rename inside the already-prefixed namespace

DetrForSegmentation                  (class registered: segmentation-only renames)
├── detr: DetrForObjectDetection     (no mapping; just walked through)
│   └── model: DetrModel             (class registered: shared base transforms)
│       └── backbone, encoder, ...
└── mask_head, bbox_attention        (head-specific weights)

"DetrModel":           [WeightRenaming("backbone.conv_encoder", "backbone"), ...]      # shared base
"DetrForSegmentation": [WeightRenaming("mask_head.lay1", "mask_head.conv1.conv"), ...] # head-specific

"ernie4_5_vl_moe": [
    WeightRenaming("vision_model", "vision_tower"),
    WeightConverter(
        source_patterns=["experts.*.down_proj.weight"],
        target_patterns=[
            "text_moe.experts.down_proj",
            "vision_moe.experts.down_proj",
        ],
        operations=[ErnieFuseAndSplitTextVisionExperts(stack_dim=0, concat_dim=1)],
    ),
],

_MODEL_TO_CONVERSION_PATTERN = {
    "mixtral": "mixtral",
    "minimax": "mixtral",
    "qwen2_moe": "qwen2_moe",
    "deepseek_v2": "qwen2_moe",
    "deepseek_v3": "qwen2_moe",
    "qwen3_moe": "qwen2_moe",
    "olmoe": "qwen2_moe",
    ...
}