laguna-xs-2.1

A tt-model package: a self-contained Docker image serving poolside/Laguna-XS-2.1 on Tenstorrent blackhole hardware.

tt-model pull  tt-hous/laguna-xs-2.1     # image -> docker, weights -> host HF cache
tt-model serve tt-hous/laguna-xs-2.1     # docker run with the right devices and mounts

Serve profiles

profile	hardware	mesh	max seqs	context	description
p150x2 (default)	p150x2	P150x2	1	131072	Qualified production default on two P150 ASICs, with canonical prefix caching.
p150x4	p150x4	P150x4	8	131072	Four-P150 regression profile with prefix caching explicitly disabled.

Select one with tt-model serve tt-hous/laguna-xs-2.1 --profile <name>; the default is p150x2.

Provenance

• type: vllm
• weights: poolside/Laguna-XS-2.1 — downloaded to the host HF cache at pull; never inside the image
• tt metal: {'sha': '3541217838f5e362623a8a385c1b8dba3fdc7e0f', 'describe': 'v0.74.0-dev20260622-222-g3541217838-dirty', 'dirty': True, 'scm_version': '0.65.2.dev7120+g3541217838f', 'mode': 'local'}
• plugin: {'repo': 'https://github.com/tenstorrent/vllm-tt-plugin', 'sha': 'c127c17d80d66ee83d23064d3a62ac844a1170de'}
• image: tt-model/laguna-xs-2.1:354121783
• tt model version: 0.2.0
• created at: 2026-08-25T02:07:22+00:00

What ships

code/ holds exactly the model's own code — the allowlist from the manifest, byte-identical to what is inside the image (both are staged from the same directory):

models/
models/autoports/
models/autoports/poolside_laguna_xs_2_1/
models/autoports/poolside_laguna_xs_2_1/doc/
models/autoports/poolside_laguna_xs_2_1/doc/datatype_sweep/
models/autoports/poolside_laguna_xs_2_1/tests/
models/autoports/poolside_laguna_xs_2_1/tests/laguna_weights.py
models/autoports/poolside_laguna_xs_2_1/tt/
models/autoports/poolside_laguna_xs_2_1/tt/__init__.py
models/autoports/poolside_laguna_xs_2_1/tt/dflash_reference.py
models/autoports/poolside_laguna_xs_2_1/tt/generator.py
models/autoports/poolside_laguna_xs_2_1/tt/generator_vllm.py
models/autoports/poolside_laguna_xs_2_1/tt/kv_grouping.py
models/autoports/poolside_laguna_xs_2_1/tt/model.py
models/autoports/poolside_laguna_xs_2_1/tt/multichip_decoder.py
models/autoports/poolside_laguna_xs_2_1/tt/optimized_decoder.py
models/autoports/poolside_laguna_xs_2_1/tt/prefill_page_table.py
models/autoports/poolside_laguna_xs_2_1/tt/prefill_runtime.py
models/autoports/poolside_laguna_xs_2_1/vllm_ext/
models/autoports/poolside_laguna_xs_2_1/vllm_ext/extra_models/
models/autoports/poolside_laguna_xs_2_1/vllm_ext/extra_models/laguna/
models/autoports/poolside_laguna_xs_2_1/vllm_ext/laguna_vllm_ext/
models/autoports/poolside_laguna_xs_2_1/vllm_ext/laguna_vllm_ext/__init__.py
models/autoports/poolside_laguna_xs_2_1/vllm_ext/laguna_vllm_ext/hybrid_kv.py
models/autoports/poolside_laguna_xs_2_1/vllm_ext/laguna_vllm_ext/lifecycle.py
models/autoports/poolside_laguna_xs_2_1/vllm_ext/laguna_vllm_ext/prefix_cache.py
models/autoports/poolside_laguna_xs_2_1/vllm_ext/laguna_vllm_ext/prefix_cache_quantum.py
models/common/
models/common/lightweightmodule.py
models/common/modules/
models/common/modules/lazy_buffer.py
models/common/modules/sampling/
models/common/modules/sampling/sampling_1d.py
models/common/modules/tt_ccl.py
models/common/readiness_check/
models/common/readiness_check/__init__.py
models/common/readiness_check/contract.py
models/common/readiness_check/contract_vllm.py
models/common/readiness_check/generate.py
models/common/readiness_check/schema.py
models/common/readiness_check/teacher_forcing.py
models/common/sampling/
models/common/sampling/__init__.py
models/common/sampling/_utils.py
models/common/sampling/generator.py
models/common/sampling/tt_log_probs.py
models/common/sampling/tt_penalties.py
models/common/sampling/tt_sampling.py
models/common/utils.py

image/ is the Docker image as an exploded OCI layout; tt-model.yaml is the full pinned manifest.