[Feature] Support MiniMax-H3 FlashGen native LoRA with the legacy manager (#6666)

Signed-off-by: Zijun Huang <zijunhua@usc.edu>
Co-authored-by: WeiQing Chen <40507679+david6666666@users.noreply.github.com>
This commit is contained in:
Zijun HuangandWeiQing Chen authored and GitHub committed 2026-08-29 06:52:19 +00:00
1 parent 143ccc5a2f
commit 08cdee8dc4
8 files changed
+1389 -6

No files matched your search

+57
View File
@@ -860,6 +860,63 @@ fixed HBM usage. Model-level and standard layerwise offload remain unsupported.
The five requested sigma points produce the four denoiser evaluations expected
by the Turbo artifact.
### FlashGen native LoRA
The FlashGen 4-step T2VA artifact uses the native MiniMax-H3 module layout and
declares its distilled sigma schedule in safetensors metadata. It is published on
[ModelScope](https://modelscope.cn/models/FlashGen/Minimax-H3-4step-lora-flashgen):
```text
FlashGen/Minimax-H3-4step-lora-flashgen/minimax_h3_t2va_flashgen_4step_v1.0_768p_bf16.safetensors
```
Download only that file:
```bash
python -m pip install modelscope
export FLASHGEN_DIR=/path/to/minimax-h3-flashgen-lora
export FLASHGEN_FILE=minimax_h3_t2va_flashgen_4step_v1.0_768p_bf16.safetensors
modelscope download FlashGen/Minimax-H3-4step-lora-flashgen \
--local_dir "${FLASHGEN_DIR}" \
--include "${FLASHGEN_FILE}"
export FLASHGEN_LORA="${FLASHGEN_DIR}/${FLASHGEN_FILE}"
```
Start from a non-offloaded or DLO FL2VA server command and add
`--task-type fl2va --lora-backend peft --lora-path "${FLASHGEN_LORA}"`.
Each request must use T2VA and the distilled interval-count contract:
```bash
-F 'num_inference_steps=4' \
-F 'extra_params={"task":"t2va","duration":5.2}' \
-F "lora={\"name\":\"h3-flashgen-v1.0\",\"path\":\"${FLASHGEN_LORA}\",\"scale\":1.0}"
```
This path rejects Ref2VA and checkpoints that already pin `base_schedule` in
`model_index.json`. The adapter metadata carries
`base_schedule=1.0,0.7,0.4,0.15,0.0`, so `num_inference_steps=4` means four
denoiser evaluations, not five sigma points. Request-mode generation may omit
the field and take the count from the adapter schedule; `--step-execution`
requires it explicitly, because the step scheduler reads the total step count
off the request at admission, before the adapter schedule is known.
DLO is supported in request-mode generation on the same terms as the Turbo
adapter: the request-switchable LoRA A/B buffers stay resident on the
accelerator while DLO streams only the base blocks, so budget for that
additional fixed HBM usage. The native artifact is rank 64 over 259 target
modules, and its packed `qkv_proj` and `fc1` layers reuse the full-input A
tensor per slice while B carries slice-local output rows, so the resident
footprint exceeds the on-disk payload; measure it for your parallel layout
rather than assuming the checkpoint size. Pure Ulysses replicates the adapter
on every rank, while DiT tensor parallelism shards the B buffers. Model-level
and standard layerwise offload remain unsupported, and `--step-execution`
cannot be combined with `--enable-distributed-layerwise-offload`.
To validate a deployment, post the same fixed-seed T2VA request twice with the
adapter and twice without it, then compare the four output digests. The adapter
is bound and deterministic when each pair matches internally and the two pairs
differ from each other.
## Key parameters
| Parameter | Recommended value | Notes |