minke.yu/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 12 Packages Projects Releases Wiki Activity

[diffusion] feat: avoid direct GPU parameter copies (#36832)

Browse Source
This commit is contained in:
Mick
2026-08-29 22:43:25 +08:00
committed by GitHub
parent cdbfe90b4a
commit b24bd44556
4 changed files with 91 additions and 26 deletions
Download Patch File Download Diff File Expand all files Collapse all files
docs/docs/sglang-diffusion/api/cli.mdx
+1 -1
View File
@@ -86,7 +86,7 @@ Use `sglang generate --help` and `sglang serve --help` for the full argument lis
- `--lora-merge-mode {auto|merge|dynamic}`: choose how LoRA is applied. `auto` statically merges regular weights and uses dynamic LoRA for FSDP-sharded weights to avoid full-gather peaks.
- `--num-gpus {N}`: number of GPUs to use
- `--performance-mode {manual|auto|speed|memory}` / `--mode`: preset for latency/throughput and memory defaults. `auto` is the default and dispatches residency from selected-GPU headroom and workload type: image DiTs stay resident above the 45 GiB threshold, while video DiT placement remains model-specific. It uses FSDP only for validated DiT-offload replacement paths. `speed` keeps `torch.compile` disabled unless a model-specific deployment config opts in after validation; pass `--enable-torch-compile true` to enable it explicitly. Use `manual` to keep performance-related server args under explicit user control. Explicit offload, FSDP, and parallelism flags take precedence in all modes.
- `--direct-gpu-weight-loading {true|false}`: opt into direct GPU loading for an unquantized, GPU-resident, TP=1 DiT by materializing its complete checkpoint state dict on GPU. Startup impact is model-dependent, so benchmark the target model before deployment. Disabled by default because checkpoint and model weights coexist temporarily, substantially increasing peak GPU memory. It is incompatible with DiT CPU/layerwise offload and FSDP.
- `--direct-gpu-weight-loading {true|false}`: opt into direct GPU loading for an unquantized, GPU-resident, TP=1 DiT by materializing its complete checkpoint state dict on GPU. Compatible tensors become parameter storage without another GPU copy, while tensors requiring transforms can still need temporary allocations. It is disabled by default; startup time and peak GPU memory remain model-dependent, so benchmark the target model before deployment. It is incompatible with DiT CPU/layerwise offload and FSDP.
- `--tp-size {N}`: tensor parallelism size. Depending on the pipeline, it can shard the DiT, one or more encoders, or both.
- `--sp-degree {N}`: sequence parallelism size
- `--dp-size {N}` (alias `--data-parallel-size`): number of data-parallel replicas. Each replica is a full copy of the engine on `num_gpus / N` GPUs with its own ingress; generation requests round-robin across replicas, realtime sessions stick to the replica holding their state, and control operations (weights, LoRA, memory occupation, shutdown) apply to every replica. Combines with the other parallelism axes (`num_gpus = dp × cfg × tp × sp`); monolithic serving only.