Skip to content

Add Turing safe-FP16 execution path #2216

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
Ownistic wants to merge 2 commits into
base: main
Choose a base branch
from
Open

Add Turing safe-FP16 execution path #2216

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
8a102dc
fix(minimax-h3): route Qwen prompt encoding through SDPA
Ownistic Aug 24, 2026
6ab55e7
fix(minimax-h3): add Turing safe FP16 execution
Ownistic Aug 24, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
11 changes: 10 additions & 1 deletion models/minimax_h3/pipeline.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -13,6 +13,7 @@
from tqdm import tqdm

from mmgp import offload
from shared.attention import attention_config_shared_state
from shared.utils.loras_mutipliers import update_loras_slists
from shared.utils.text_encoder_cache import TextEncoderCache
from shared.utils.frame_scheduler import floor_frame_count, normalize_frame_count, normalize_overlap
Expand Down Expand Up @@ -290,6 +291,7 @@ def __init__(self, transformer, text_encoder, video_vae, audio_vae, latent_upsca
self.text_encoder_cache = TextEncoderCache()
self._shared_offloadobj = None
self._private_offloadobj = None
self._attention_split_logged = False
self._interrupt = False

def set_offload_handoff(self, shared_offloadobj, private_offloadobj):
Expand Down Expand Up @@ -371,7 +373,14 @@ def _prompt_cache_key(self, prompt, presentation):

def _encode_prompt(self, prompt, presentation):
def encode_fn(prompts):
return [self.text_encoder.encode(prompts[0], presentation, self.device, self.dtype)]
# Qwen3-VL's Sage kernel is not supported on Turing, while H3 can still
# use the selected backend after this context restores it.
if not self._attention_split_logged:
transformer_attention = offload.shared_state.get("_attention", "auto")
print(f"[MiniMax H3] Qwen attention=SDPA; transformer attention={transformer_attention}")
self._attention_split_logged = True
with attention_config_shared_state("sdpa"):
return [self.text_encoder.encode(prompts[0], presentation, self.device, self.dtype)]

cache_key = self._prompt_cache_key(prompt, presentation)
return self.text_encoder_cache.encode(encode_fn, prompt, device=self.device, cache_keys=cache_key)[0]
Expand Down
Loading