Distributed Layerwise Offloading¶

Distributed layerwise offloading (DLO) extends block streaming to multi-device deployments. Each selected component chooses its transfer independently. With allgather, each rank stores roughly one shard per member of the effective DLO group (DP, or SP when DP is one) and reconstructs each layer at runtime. With rank-local, each rank streams a complete loader-produced block independently. Compatible TP1 DiT deployments can share checkpoint-backed host pages among processes on the same node; otherwise DLO streams the ordinary loader's rank-local tensors.

See the DLO feature design for the implementation contract and compatibility matrix.

Execution model¶

DLO overlaps three operations with a fixed two-block device buffer:

Compute stream:  [Layer N]          [Layer N+1]        [Layer N+2]
H2D stream:      [H2D shard N+1]    [H2D shard N+2]
AllGather:       [AG N+1]           [AG N+2]
Slots:           slot 0: Layer N    slot 1: Layer N+1

AllGather communicates only request-independent weight shards, so data- parallel ranks may process different requests concurrently.

The public API describes intent in one nested configuration:

weight_transfer applies only to model weights. It does not change encoder activations or the encoder's compute parallelism. rank-local copies each rank's loader-produced block to its device. allgather stores one shard of that block per rank, copies each shard to its device, and reconstructs the complete block with a device-side AllGather before compute.

The historical enable_cpu_offload, enable_layerwise_offload, and enable_distributed_layerwise_offload fields remain compatibility aliases for existing callers and model-specific stage lifecycles. Existing combinations retain their priority (distributed layerwise > layerwise > module). A compact config rejects conflicting compatibility strategies and non-default legacy DLO tuning. distributed-layerwise remains an internal backend name; users select mode="layer", and the runtime chooses the capable backend.

Usage¶

# Four SP ranks; Wan declares replicated text-encoder weights, so both
# selected components can use AllGather safely.
vllm serve Wan-AI/Wan2.2-T2V-A14B-Diffusers --omni \
  --diffusion-offload-config \
  '{"mode":"layer","components":["dit","text_encoder"],"layer_options":{"dit":{"weight_transfer":"allgather"},"text_encoder":{"weight_transfer":"allgather"}}}' \
  --usp 4

# Mix transfers independently: sharded DiT, loader-produced encoder layout
vllm serve /path/to/model --omni \
  --diffusion-offload-config \
  '{"mode":"layer","components":["dit","text_encoder"],"layer_options":{"dit":{"weight_transfer":"allgather"},"text_encoder":{"weight_transfer":"rank-local"}}}' \
  --usp 4

# Full-topology rank-local DLO for an existing model-specific lifecycle
vllm serve /path/to/model --omni \
  --enable-distributed-layerwise-offload \
  --dlo-no-use-allgather \
  --usp 4

A compact layer config whose selected components are all rank-local and whose resident_layers is zero resolves to the ordinary layerwise backend. It uses the ordinary loader and does not request DLO's direct-checkpoint mmap. Keep the compatibility DLO flags for a model-specific full-topology rank-local lifecycle; do not add resident layers solely to force backend selection.

from vllm_omni import Omni

omni = Omni(
    model="/path/to/model",
    diffusion_offload_config={
        "mode": "layer",
        "components": ["dit", "text_encoder"],
        "layer_options": {
            "dit": {
                "weight_transfer": "rank-local",
                "resident_layers": 20,
            },
            "text_encoder": {
                "weight_transfer": "allgather",
            },
        },
        "pin_memory": True,
    },
)

Flags¶

HWR and host-registration tuning remain on the compatibility DLO interface and cannot be combined with diffusion_offload_config in this release.

Component and weight-transfer matrix¶

The text encoder's compute group and weight-transfer group are separate concepts. An encoder TP group owns different parameter shards, so it is not a valid AllGather offload group. Models opt in through OffloadPlan.encoder_dlo_weight_replication; otherwise multi-rank encoder AllGather fails at startup with guidance to use rank-local. This makes SP AllGather available to replicated encoders without silently corrupting an encoder-TP layout.

resident_layers remains a DiT-only setting and currently requires the DiT transfer to be rank-local.

Host-weight loading¶

When the distributed backend is selected by AllGather, resident DiT layers, or the compatibility DLO flag, the diffusion loader chooses host storage before DLO is enabled. It first attempts to build a complete, validated direct-checkpoint mmap plan. If names, coverage, shape, dtype, topology, or loader-callback compatibility cannot be proven, it runs the ordinary model loader instead. DLO consumes that result and does not make a second checkpoint-compatibility decision.

The shared-mmap optimization in this phase is supported only with TP1. TP greater than one falls back before model mutation to ordinary TP-aware loading. DLO may still consume those TP-local tensors, but this is a compatibility path: it does not share checkpoint-backed runtime weights across DP replicas and provides no shared-mmap host-memory guarantee.

The mmap plan skips only dedicated DiT weight sources. Other component sources, such as a text encoder loaded through the shared diffusion loader, continue to use their ordinary component loader. A checkpoint source that mixes DiT and non-DiT weights falls back completely rather than leaving an unplanned component uninitialized.

With direct checkpoint mmap, the loader:

1. saves non-persistent buffers such as RoPE frequencies;
2. moves the normally created transformer to the meta device;
3. loads checkpoint tensors as mmap views backed by the shared OS page cache;
4. applies any loader-owned bounded layout adapters while packing blocks;
5. restores saved non-persistent buffers; and
6. preserves post_load_weights() and validate_loaded_weights() lifecycle hooks.

For a DiT using allgather with a group larger than one, each process copies only its persistent shard and then releases the source mapping. For a DiT using rank-local, each process keeps the mapping open and packs complete blocks through two bounded pinned staging slots. Processes mapping the same files on one node share the immutable pages; rank-local transfer still performs a complete-block H2D copy in each process.

When the effective DLO group size is one, weight_transfer="allgather" does not perform a collective and uses the same rank-local transfer behavior.

Final-layout Host Weight Runtime¶

HWR is an opt-in startup optimization for models that declare the final-layout BF16 restore contract. Use it only when the selected DiT uses rank-local:

The validated BF16 model contracts currently cover MiniMax H3 and black-forest-labs/FLUX.2-klein-4B. FLUX.2-klein-9B shares the same model class but has not been validated against this contract. FLUX.2-dev, online FP8, HSDP, LoRA/adapted weights, and non-default load formats remain outside the validated scope.

vllm serve /path/to/model --omni \
  --enable-distributed-layerwise-offload \
  --dlo-no-use-allgather \
  --host-weight-runtime-mode preferred \
  --host-weight-runtime-root /var/cache/vllm-omni/hwr

Choosing a mode¶

Both enabled modes still fail on non-retryable configuration, identity, or compatibility errors. preferred is a fallback policy for a cache miss or a recoverable cache problem; it does not hide an invalid deployment.

Populating a store for required mode¶

required is deliberately consume-only. PR2 does not include a separate prewarm command, so populate each node-local storage domain with one matching preferred producer cohort:

1. Choose a persistent, writable root visible to every diffusion worker in the storage domain. Do not use a process-private temporary directory.
2. Start the deployment with preferred, using the exact model revision, dtype, TP size, SP configuration, and other weight-layout settings intended for serving. No inference request is needed; publication happens during model startup. Wait for the engine to become healthy, then shut it down cleanly.
3. Restart with the same arguments and root, changing only the mode to required. A successful startup proves that every worker acquired a valid artifact; a miss, corrupt artifact, or incompatible identity fails startup.
4. Repeat the population start on every node or storage domain because this store is node-local.

For example:

# First startup: canonically load and populate the exact artifacts.
vllm serve /path/to/model --omni \
  --enable-distributed-layerwise-offload \
  --dlo-no-use-allgather \
  --host-weight-runtime-mode preferred \
  --host-weight-runtime-root /var/cache/vllm-omni/hwr

# After a healthy startup and clean shutdown, enforce cache hits.
vllm serve /path/to/model --omni \
  --enable-distributed-layerwise-offload \
  --dlo-no-use-allgather \
  --host-weight-runtime-mode required \
  --host-weight-runtime-root /var/cache/vllm-omni/hwr

Include the same TP and SP layout flags in both commands. DP rank and DP size are excluded from artifact identity, so the population and serving DP sizes may differ and equivalent DP replicas share artifacts. TP rank is included, so a TP-N deployment normally needs N rank-specific artifacts in each storage domain; launching the matching TP cohort creates that set. If the model revision or layout changes, run preferred again for the new identity before returning to required.

On a cold start, the canonical loader remains authoritative and publishes a validated final-layout artifact for later workers. A warm start restores the DiT final tensors without ordinary DiT materialization. DLO then attempts to register the immutable mapping for direct H2D; if registration is unavailable, it streams through the same two bounded host staging slots. The artifact identity includes the TP rank/size and SP layout, so TP1, TP2 rank-local shards, and distinct SP layouts do not alias one another. Publication failure remains separate from a preferred serving startup; the next required startup provides the explicit artifact-availability check.

For local canonical checkpoints, the first eligible worker may hash source shards to establish immutable identity. HWR caches those digests in the same node-local storage domain and validates file metadata before reuse, so later workers normally avoid repeating that read. Cold BF16 publication also hashes ordered payloads as they are written and overlaps payload durability work with later shards; this changes startup work only, not artifact contents or runtime H2D behavior.

Registered direct H2D¶

Registration is attempted automatically only for an eligible warm HWR hit when the existing pin_cpu_memory policy is enabled. A successful path is:

shared read-only HWR mmap -> existing rotating HBM block buffer -> GPU kernel

It removes the recurrent mmap-to-private-staging CPU copy and does not allocate the two host staging slots. It does not reduce the H2D payload or make GPU kernels access host memory directly.

CUDA requires read-only host-registration capability for the immutable HWR mapping. Unsupported capability, a positive registration limit smaller than the complete page-aligned mapping, or a safely rolled-back registration error falls back to bounded staging. Programmatic compatibility configurations can set pin_cpu_memory=False to disable registration explicitly. A successful registration locks the complete mapped range in host memory for that worker's lifetime, so use --dlo-host-registration-limit-gib when an operator-enforced ceiling is required.

Shutdown drains H2D work, releases hook/source references, unregisters every range, and only then closes the HWR lease. AllGather and direct checkpoint mmap retain their existing transfer paths.

When HWR is disabled, DLO is disabled, or the DiT uses allgather, the loader does not resolve HWR sources or construct its store.

Declarative topology¶

Models may declare an OffloadPlan instead of embedding offload logic:

from vllm_omni.diffusion.offloader import OffloadPlan


class MyPipeline(nn.Module):
    _dit_modules = ["transformer"]
    _encoder_modules = ["prompt_model"]
    _offload_plan = OffloadPlan(
        block_attrs={"transformer": ("blocks",)},
        offload_submodules={"context_encoder": "layers"},
        encoder_component_types={"prompt_model": "text_encoder"},
        encoder_block_attrs={"prompt_model": ("encoder.layers",)},
        encoder_dlo_weight_replication=frozenset({"prompt_model"}),
        on_demand_component_paths=frozenset({"prompt_model"}),
    )

encoder_component_types maps arbitrary encoder paths to the public text_encoder selector. encoder_dlo_weight_replication is the explicit safety declaration for reusing the DiT DLO group. On-demand encoder non-block state is loaded and offloaded by its pipeline phase.

When no plan exists, DiT discovery falls back to _layerwise_offload_blocks_attrs and then heuristic attribute lookup; undeclared auxiliary components remain resident.

Data-parallel concurrency¶

With data_parallel_size > 1 and AllGather enabled, the scheduler can process up to dp_size requests per denoising step. Concurrent requests must resolve to the same denoise schedule, so every request must provide the same explicit num_inference_steps. Pipeline defaults are rejected because different request modes can resolve the same None value to different schedules.

Limitations¶

• Direct checkpoint mmap currently requires TP1. TP greater than one is outside the Phase A shared-mmap support scope and falls back before model mutation to the ordinary TP-aware loader. DLO can stream that runtime layout, and eligible DiT rank-local configurations may use HWR final-layout artifacts, but direct checkpoint mmap provides no shared-mmap host-memory guarantee for TP greater than one.
• HSDP with allgather on any selected component is rejected to avoid double sharding. HSDP with rank-local transfers has limited end-to-end validation.
• With data parallelism, allgather cannot be combined with TeaCache or Cache-DiT on any selected component. Prompt-embedding cache is also incompatible when the text encoder uses allgather. Rank-local cache hits could otherwise make ranks enter different weight collectives.
• Per-tensor online FP8, INT8, and MXFP8 linears use the ordinary loader and can run with either DiT transfer path. With DiT allgather, every rank temporarily materializes the complete quantized model in host memory before DLO retains only its shard. Layers too wide for the NPU quantization kernel (npu_quant_matmul rejects an output dimension past 65535) stay unquantized and load straight into host memory under offload-after-quant — their runtime layout is a plain contiguous bf16 weight, the same layout DLO shards on the ordinary path, so they are allowed under allgather too. Other online quantization methods require rank-local transfer for the affected component until their runtime layouts are validated.
• Resident leading layers require DiT rank-local transfer and a model OffloadPlan that declares eligible resident_dit_paths.
• DP concurrency requires an explicit, identical inference-step count.

Sharing quantized or otherwise unvalidated transformed layouts through a normalized HWR producer is a follow-up design in RFC #6195, not part of the current BF16 final-layout path.

See the Cosmos3 DistOffload recipe for an end-to-end example.