Transformers documentation
Dynamic weight loading
Dynamic weight loading
Checkpoints are often serialized in a format that does not match what a model expects at runtime. Common scenarios include:
- Fused weights: Checkpoints store separate
gate_projandup_projweights, but the model uses a fusedgate_up_projfor efficiency. - MoE expert consolidation: Individual expert weights (
experts.0.weight,experts.1.weight, …) need to be stacked into a single 3D tensor. - Legacy naming: Old checkpoints use different naming conventions (e.g.,
LayerNorm.gammavsLayerNorm.weight). - Quantization: Weights may be stored in quantized formats that need deserialization.
Dynamic weight loading addresses this by applying scheduled, reversible operations to checkpoint tensors as they are loaded. Transformers makes this available through WeightConverter, which maps one or more source keys to target keys by running a list of composable conversion operations. This approach adapts to new weight layouts, and supports loading quantized mixture-of-experts (MoEs) or enabling tensor parallelism and MoEs.
This guide demonstrates how to use the WeightConverter to convert tensors. Your WeightConverter should be added inside _build_checkpoint_conversion_mapping() in the conversion_mapping.py file.
Full loading pipeline
All models go through the dynamic weight loading system. Conversion mapping is an optional step within that system that only activates when the model has entries in _MODEL_TO_CONVERSION_PATTERN.
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 │
└───────────────────────────────────────────────────────────┘| Step | When it activates |
|---|---|
| Dynamic loading | Always, for all models |
| Conversion mapping | Only when model_type is in _MODEL_TO_CONVERSION_PATTERN |
| TP sharding | Only when tp_plan="auto" and model has base_model_tp_plan |
| Dequantization/deserialization | Only when loading a pre-quantized checkpoint |
| Quantization | Only when a quantization config is provided and weights are not pre-quantized |
Dense models (e.g., Llama)
For most dense models, the checkpoint format matches the model format directly, so no conversion mapping is needed. Some models may still require renaming (e.g., legacy naming conventions). TP sharding still applies when enabled.
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
```x
Legacy checkpoints may use older naming conventions that are handled by built-in renamings applied to all models:
Checkpoint: Model: LayerNorm.gamma → LayerNorm.weight LayerNorm.beta → LayerNorm.bias
### MoE models (e.g., Mixtral)
For MoE models, the checkpoint format differs from the model format. Conversion mapping transforms separate expert weights into fused 3D tensors, and TP sharding applies after conversion.
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 ─┘
## Architecture
The system is built around several key components defined in `src/transformers/core_model_loading.py`:
**Phase 1 — Per-key processing** (iterates over checkpoint keys):
1. **Rename key** via `WeightRenaming` (e.g. `block_sparse_moe` -> `mlp`)
2. **Match pattern** via `WeightConverter` (e.g. `experts.*.w1.weight`)
3. **Shard (TP) and send to device** asynchronously via `ThreadPoolExecutor`
4. **Collect** tensors with the same `source_pattern` together (e.g. all MoE expert weights, gate + up projections)
**Phase 2 — Per-mapping processing** (iterates over collected mappings):
1. **Dequantize/deserialize** (pre-quantized checkpoints only)
2. **Apply `ConversionOps` chain**: `Chunk`, `Concatenate`, `MergeModulelist`, `Transpose`, etc.
3. **Quantize** on-the-fly (if not pre-quantized)
4. **Set parameter** on model
### WeightTransform
The base class that handles pattern matching and tensor collection:
- **Pattern compilation**: Converts glob-style patterns (`*.weight`) to regex.
- **Key renaming**: `rename_source_key()` transforms checkpoint keys to model keys.
- **Tensor collection**: `add_tensor()` gathers related tensors for batch processing.
- **Reversibility**: `reverse_transform()` creates the inverse operation for saving.
```python
@dataclass(slots=True)
class WeightTransform:
source_patterns: str | list[str] # Checkpoint key patterns
target_patterns: str | list[str] # Model key patterns
compiled_sources: re.Pattern # Compiled regex for matching
distributed_operation: TensorParallelLayer | None
quantization_operation: ConversionOps | None
collected_tensors: dict[str, list[Future]] # Gathered tensors
layer_targets: dict[str, set[str]] # Target key trackingWeightRenaming
WeightRenaming is a specialized WeightTransform for simple 1:1 key renaming without tensor operations:
# Legacy checkpoint compatibility
WeightRenaming("LayerNorm.gamma", "LayerNorm.weight")
# Module path changes
WeightRenaming(".block_sparse_moe.", ".mlp.")
# Adding prefixes
WeightRenaming("(.+)", "timm_model.\\1")WeightConverter
WeightConverter extends WeightTransform with a list of ConversionOps:
@dataclass(slots=True)
class WeightConverter(WeightTransform):
operations: list[ConversionOps] # Chain of operationsIt supports many-to-one (e.g., concatenating gate + up → gate_up), one-to-many (e.g., splitting qkv → q, k, v), and chained operations applied sequentially.
Conversion operations
The WeightConverter class has several operations that are executed when from_pretrained() is called for transforming checkpoint source tensors into model target tensors.
Operations are fully reversible. Saving reverses the conversions and returns the original checkpoint so you can easily work across different frameworks.
| Operation | Reverse |
|---|---|
Chunk(dim) | Concatenate(dim) |
Concatenate(dim) | Chunk(dim) |
MergeModulelist(dim) | SplitModulelist(dim) |
SplitModulelist(dim) | MergeModulelist(dim) |
Transpose(d0, d1) | Transpose(d1, d0) |
Force16BytesAlignment | Force16BytesAlignment (idempotent) |
Chunk
The Chunk operation splits a tensor into equal parts along a dimension. For example, if a model expects Q, K, and V as three separate tensors instead of a single tensor.
WeightConverter(
"self_attn.qkv_proj",
["self_attn.q_proj", "self_attn.k_proj", "self_attn.v_proj"],
operations=[Chunk(dim=0)],
)Concatenate
The Concatenate operation fuses separate tensors into a single tensor. For example, if a model expects Q, K, and V as a single tensor instead of separate tensors.
WeightConverter(
["self_attn.q_proj", "self_attn.k_proj", "self_attn.v_proj"],
"self_attn.qkv_proj",
operations=[Concatenate(dim=0)],
)MergeModulelist
MergeModulelist merges a list of 2D tensors into a single 3D tensor. For example, you can compose MergeModulelist with Concatenate to stack the experts in a MoE and pack them into one tensor.
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),
],
)SplitModulelist
SplitModulelist splits a 3D tensor back into a list of 2D tensors. For example, you can split a stack of experts back into individual experts.
WeightConverter(
"mlp.experts.down_proj",
"block_sparse_moe.experts.*.w2.weight",
operations=[SplitModulelist(dim=0)],
)PermuteForRope
PermuteForRope converts weights from the interleaved format to use the sin/cos format. For example, you can compose Chunk with PermuteForRope to split a fused QKV tensor and apply the sin/cos RoPE permutation to Q and K.
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(),
],
)Transpose
Transpose swaps dimensions of a tensor. Useful for converting weight layouts between different conventions.
WeightConverter(
source_patterns="mlp.gate.weight",
target_patterns="mlp.text_moe.gate.weight",
operations=[Transpose(dim0=0, dim1=1)],
)Force16BytesAlignment
Force16BytesAlignment clones a tensor if it is not 16-byte aligned. This is required for torch._grouped_mm and TMA/SIMD operations. It is idempotent: applying it more than once has no additional effect.
Operation chaining
Operations can be chained to perform complex transformations. The operations execute in order, with each operation’s output becoming the next operation’s input.
Example: Mixtral MoE conversion
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)
],
)Data flow:
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_upPattern matching
The * in patterns acts as a wildcard:
- During loading, it matches any numeric index (
experts.0.,experts.1., etc.). - Tensors with the same pattern (differing only in index) are grouped together.
- The order of collection is preserved for correct concatenation.
Tensor parallelism integration
The dynamic loading system integrates with tensor parallelism (TP) through the TensorParallelLayer hierarchy defined in src/transformers/integrations/tensor_parallel.py.
When TP is enabled, tensors are sharded during materialization, not after. This means each rank only loads the portion of the tensor it needs.
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)Available parallel styles
| Style | Weight Shard Dim | Description |
|---|---|---|
colwise | -2 | Column-wise: output features sharded |
rowwise | -1 | Row-wise: input features sharded |
packed_colwise | -2 | For fused weights (gate_up_proj) |
packed_rowwise | -1 | For fused weights |
embedding_rowwise | 0 | Vocabulary parallelism |
grouped_gemm | 0 | Expert parallelism for MoE |
sequence_parallel | None | No weight sharding |
Packed weight handling
For fused weights like gate_up_proj, special care is needed to shard correctly:
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]
"""The TP operation is stored in the WeightTransform and applied after conversion operations:
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()
)Quantization integration
Quantization hooks into the loading pipeline in two ways, depending on whether the checkpoint is already quantized:
- Pre-quantized checkpoints: The quantizer provides WeightConverter instances (via
get_weight_conversions()) that deserialize quantized tensors. Checkpoint dtypes are preserved to avoid unwanted casts. - On-the-fly quantization: The quantizer provides a quantization operation that is applied after conversion ops, quantizing weights as they are loaded.
Fast and efficient model loading
Loading a model is faster and uses less memory because the loader knows which tensors are required for operations and schedules their materialization lazily.
The loader scans the checkpoint once to discover pattern matches and collect tensors. It stores them as Future objects and submits them to a thread pool for asynchronous loading without blocking the GIL. A parameter starts loading as soon as a thread becomes available to it.
If your system runs other heavy processes, multiple threads may slow down loading instead of accelerating it. In this case, set the environment variable HF_DEACTIVATE_ASYNC_LOAD=1 to load weights sequentially.
The default is 4 threads for asynchronous parameter loading. This provides the best trade-off across loading scenarios and hardware. The work is mostly I/O bound, but depending on accelerator hardware and the
dtyperequired at loading, it can become CPU/GPU-bound if thedtypediffers from the serialized one (this requires an additional copy operation).
Async vs sync loading
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)Sync loading is used when:
HF_DEACTIVATE_ASYNC_LOAD=1environment variable is set.- Disk offloading is enabled (memory constraints require sequential loading).
Materialization flow
1. Checkpoint iteration (Phase 1):
- For each key, 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)Memory efficiency
When converting a weight, the converter waits for all required tensors to materialize if they haven’t loaded yet. For example, the MergeModulelist operation requires all weights in ModuleList to be loaded before merging.
Concatenating tensors requires a temporary copy, so operations like MergeModulelist and Concatenate need 2x the memory of the underlying tensors during conversion. Once merged, only the resulting tensor stays in memory. The theoretical worst-case memory peak is the model size plus the tensors required for the largest MergeModulelist or Concatenate operation.
This worst case only occurs when all other parameters have loaded before the demanding conversion runs. Two scenarios trigger this.
- All parameters loaded asynchronously before entering the demanding conversion (the thread pool was faster than the conversion queue).
- The demanding conversion is the last one.
For example, a MoE model using MergeModulelist for experts on each layer, the theoretical worst-case memory peak is model size plus experts on one layer.
These worst-case scenarios are uncommon. The actual memory peak tends to stay close to the model size.
Reversibility
The system supports saving models with the inverse transformations, enabling round-trip save/load:
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)Target patterns may contain regex elements that need processing for the reverse direction:
def process_target_pattern(pattern: str) -> tuple[str, str | None]:
"""
- Removes `^` and `$` anchors
- Removes negative lookahead/lookbehind
- Detects capturing groups, replaces with \1
"""Real examples
Mixtral-style MoE
Checkpoint format:
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.weightModel format:
model.layers.0.mlp.experts.gate_up_proj # (8, 4096, 28672)
model.layers.0.mlp.experts.down_proj # (8, 14336, 4096)Conversion mapping (from conversion_mapping.py):
"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)],
),
],Custom operations (ERNIE 4.5 VL MoE)
When the built-in operations aren’t sufficient, you can create a custom ConversionOps subclass. For example, ERNIE 4.5 VL MoE needs to split a shared expert list between text and vision modalities — something no single built-in op handles. The custom ErnieFuseAndSplitTextVisionExperts operation splits and re-stacks experts across two target keys:
"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)],
),
],Custom ops must implement convert() and the reverse_op property to support round-trip save/load.
Model type aliases
Many models share conversion patterns:
_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",
...
}Reusing the dynamic loading building blocks
Dynamic weight loading is not limited to full model checkpoints. The same building blocks let you load any set of weights as long as you can describe how checkpoint keys map to parameters and ensure the target modules exist.
At a high level, the contract looks like this:
- Prepare the model namespace. Make sure the modules/parameters you want to load are present and named the way your
mapping will target them. For adapters, that means calling
inject_adapter_in_model(...)so adapter modules exist before loading. For custom heads or extra modules, instantiate them on the model first. - Describe how to map weights. Build a conversion/renaming list (for example, in a helper like
_build_peft_weight_mapping(...)) using WeightConverter orWeightRenaming. This is where you express how checkpoint keys should be converted, split, merged, or renamed to match your model namespace. You can do mostly 3 things:- add operations to the list of converters: these will be applied on all weights except for the ones collected in any of the
WeightConverter. These in general should beWeightRenamingoperations - add operations to the list of operations of each converter: this is what happens for
Quantization, where we just add a quantization operation after the list of operations of anyWeightConverter. - replace / map operations to your custom operations: this is what happens with
peft. We replace theConcatenateoperation of saymixtral, to bePeftConcatenate. This way, when the adapter checkpoint is read, the weights to be concatenated are collected, and are properly formatted forpeft
- add operations to the list of converters: these will be applied on all weights except for the ones collected in any of the
- Load + finalize + report. Use the core loader to perform the conversion and populate tensors, then finalize and
log results. Concretely, this flow is:
LoadStateDictConfig(...)+_load_pretrained_model(...)to load and convert._finalize_load_state_dict(...)to move any missing/mismatched tensors offmeta, initialize them, and tie weights.log_state_dict_report(...)to report missing/unexpected/mismatched keys (and conversion errors).
These APIs are exposed to allow you to handle custom code, custom weight formats, but also make sure you benefit from the highest and most efficient weight loading, sharding and good quality of life of transformers API!
Key files reference
| File | Purpose |
|---|---|
src/transformers/core_model_loading.py | Core loading logic, WeightConverter, ConversionOps |
src/transformers/conversion_mapping.py | Built-in conversion patterns for all models |
src/transformers/integrations/tensor_parallel.py | TP sharding classes and utilities |
src/transformers/quantizers/base.py | Quantization hooks and base class |