Skip to content

Specifying explicit GPU split crashes Strata #644

Closed

Description

@tonydiep

ISSUE

Specifying explicit GPU split crashes Strata.

Multi-GPU Documentation says Strata supports specifying split in strata-.json but Strata crashes on any value except "auto"

#strata-<model>.json

 "layer_split": "auto"  <-------works
 "layer_split": "24,16,12,12"  <-------fails
 "layer_split": [24,16,12,12]  <-------fails

Specifying Split Fails

[strata] layer split across GPUs [2, 0, 1, 3] (24,16,12,12)
[strata] starting the engine: reading the model's weights ...
Traceback (most recent call last):
File "/home/tonydiep/Strata/serve/server.py", line 2710, in
sys.exit(main())
~~~~^^
File "/home/tonydiep/Strata/serve/server.py", line 2587, in main
engine = StrataEngine(exe, engine_args(cfg), cwd=cfg.get("cwd"), log=cfg.get("log"), env=env, lazy=lazy)
File "/home/tonydiep/Strata/serve/server.py", line 291, in init
raise RuntimeError("the engine exited before it was ready" + (f" (see {log})" if log else "") +
start_failure_hint(log, log_start))
RuntimeError: the engine exited before it was ready (see /home/tonydiep/Strata/strata-iq3_xxs.log)
tonydiep@tiny:~/Projects/optimise-llms$ tail /home/tonydiep/Strata/strata-iq3_xxs.log)
bash: syntax error near unexpected token `)'

"auto" Layer Works

[strata] layer split across GPUs [2, 0, 1, 3] (auto)

Activity

  1. Niko1221 commented on Oct 4, 2026

    @Niko1221
    Owner

    Thanks for the report. layer_split is the first layer of each later GPU, not the number of layers per card: for 4 GPUs it is 3 rising numbers, e.g. "layer_split": "24,36,42" puts layers 0-23 on the first card, 24-35 on the second, 36-41 on the third and 42-47 on the last. From 0.1.39 the server checks the value before the start and says what is wrong, with an example for your number of GPUs. It also accepts a JSON list ([24, 36, 42]), and MULTI_GPU.md explains the format. Your "24,16,12,12" as layer counts would be "24,40,52", which is past the model's 48 layers. "auto" picks the placement from each card's free VRAM, and is usually the best start.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions