pi05-base-p300x2

pi-0.5 (lerobot/pi05_base) on two Tenstorrent Blackhole p300 boards via tt-nn, in two serve profiles: single-robot (53 ms per action chunk) and multi-robot (two chip pairs, 47 action chunks per second for four robots).

Runs on p300x2 — see the serve profiles below.

Packaged and published with tt-model-manager 0.1.0 (manifest schema 5.1).

Quickstart

tt-model pull  changh95/pi05-base-p300x2 --with-weights
tt-model serve changh95/pi05-base-p300x2

pull --with-weights fetches the image and the pinned lerobot/pi05_base weights. serve listens on port 20000 and is ready when it logs Application startup complete (a few minutes on first start).

Run with tt-cli

tt serve changh95/pi05-base-p300x2                          # single-robot (default)
tt serve changh95/pi05-base-p300x2 --profile multi-robot    # several robots share the box
printf '{"images":["%s","%s"],"prompt":"pick up the cube","state":[0.1,-0.2,0.3,0,0,0,0.5,-0.5]}' \
  "$(base64 -w0 media/sample_base.png)" "$(base64 -w0 media/sample_wrist.png)" > req.json
curl -s localhost:20000/predict -H 'Content-Type: application/json' -d @req.json
tt model stop changh95/pi05-base-p300x2

API

• POST /predict
  • images: 1-2 base64 PNG/JPEG, base camera then wrist camera.
  • prompt (task text) or tokens (≤224 PaliGemma ids).
  • optional state (≤32 floats in [-1, 1]) and seed.
• Response: actions (50 × 32, normalised), batched_as, timing_ms.
• GET /health, GET /info.

single-robot profile (default)

All four chips work on every request. The SigLIP + Gemma-2B prefix is tensor-parallel over the 4 chips; the Gemma-300M action expert is replicated. One Metal trace, batch 1. Concurrent requests run one after another.

Latency vs batch (in process, traced; 2 × 224² images, 224 tokens, 10 denoising steps):

batch	prefix (4 chips)	action expert, 10 steps (4 chips)	total	per request
1	18.4 ms	33.7 ms	50.5 ms	50.5 ms
2	26.7 ms	39.8 ms	64.9 ms	32.4 ms
4	50.9 ms	65.6 ms	117.3 ms	29.3 ms

The profile serves batch 1. Rows 2 and 4 are the same layout with PI05_BATCH_SIZES=1,2,4.

Accuracy and speed:

metric	value
PCC vs the fp32 torch reference, 2 observations	0.9986 / 0.9988
PCC, 16 random observations	min 0.93, mean 0.98, median 0.99
HTTP, 1 client	53 ms per chunk
HTTP, 2 / 4 / 8 clients	19.5 req/s at 102 / 204 / 408 ms per chunk
same code on one p300 chip	84 ms per chunk
determinism	repeats bit-exact

multi-robot profile

The four chips form two independent pairs. Each pair runs the whole model: the prefix tensor-parallel over its 2 chips, the action expert replicated on both. A request goes to the pair with fewer requests in flight; requests arriving within 4 ms at a busy pair share a batch of 2. Sized for up to 4 robots.

Latency vs batch, one pair (in process, traced):

requests on the pair	prefix (2 chips)	action expert, 10 steps (2 chips)	total
1	24.6 ms	33.7 ms	57.2 ms
2	38.0 ms	42.5 ms	78.1 ms

Robots vs latency and throughput (HTTP, closed loop, median per request):

robots	per pair	latency	throughput
1	1	63 ms	15.8 req/s
2	1 + 1	64 ms	31.0 req/s
3	2 + 1	81 ms	37.5 req/s
4	2 + 2	84 ms	47.3 req/s
8	queueing	158 ms	49.2 req/s

Accuracy and speed:

metric	value
PCC vs the fp32 torch reference, 2 observations	0.9955 / 0.9879 (2-chip prefix partials)
two pairs concurrently, 1 / 2 requests each (in process)	57.9 / 82.0 ms wall
determinism	repeats bit-exact; both pairs give identical outputs

Beyond four robots, PI05_DP_GROUP=1 runs four independent single chips instead of two pairs. A chip alone takes 84 ms per chunk, but the expert is no longer duplicated, so the ceiling is higher:

robots	two pairs (this profile)	four single chips
4	85 ms, 47.1 req/s	94 ms, 42.3 req/s
6	119 ms, 45.4 req/s	95 ms, 47.4 req/s
8	160 ms, 48.5 req/s	146 ms, 51.3 req/s
16	320 ms, 48.8 req/s	273 ms, 56.9 req/s

Response

{"actions": [[-0.0203, -0.0197, -0.0349, ...], ...], "action_horizon": 50, "action_dim": 32, "normalized": true,
 "denoising_steps": 10, "num_tokens": 142, "batched_as": 1, "timing_ms": {"preprocess": 1.1, "inference": 50.5, "total": 51.6}}

actions are in lerobot's normalised QUANTILES space, zero-padded to 32 dims. Denormalise with (a+1)*(q99-q01)/2+q01 from your dataset and slice to your action dim.

Caveats

• Fixed inputs: 1-2 images at 224 × 224, ≤224 prompt tokens, ≤32 state floats, 10 steps.
• Needs all four chips of a 2× p300 box. One profile at a time. multi-robot alone costs 63 ms per chunk (2-chip prefix); beyond 4 robots requests queue.
• The tokenizer google/paligemma-3b-pt-224 is gated: accept the Gemma terms and hf auth login, or send tokens.
• Base checkpoint, normalised actions. No closed-loop benchmark was run.

Licensing

• Weights: lerobot/pi05_base under the Gemma Terms of Use, not redistributed here.
• Port and serving code (code/): Apache-2.0 headers, same Gemma terms.

Links

lerobot/pi05_base · arXiv:2504.16054 · Physical-Intelligence/openpi · single chip: changh95/pi05-base-p150 · design notes: code/models/experimental/pi0_5/README.md

Serve profiles

Pick one with --profile: single-robot (default) or multi-robot. Both run on p300x2; see the table above.

Provenance

The exact sources the image was built from — code/ in this repo is byte-identical to the model code inside the image:

component	built from
tt-metal	975015c2f03bb818eaee2422c3845fba381eaf8c
code/ digest	c9c924f989682aea (sha256, first 16 hex digits)
built	2026-09-18T21:21:27+00:00 by tt-model 0.1.0