GB10/Blackwell installability: relax torch pin, fix meta-tensor load crash, voice-prompt dtype cast

[jethac]

Motivation

PersonaPlex cannot currently be installed or loaded on NVIDIA GB10-class
machines (DGX Spark, ThinkStation PGX, Jetson Thor — sm_121, aarch64):

1. moshi/pyproject.toml pins torch < 2.5, but sm_121 requires the CUDA
   13.0 aarch64 wheels (torch >= 2.9, download.pytorch.org/whl/cu130).
   Community GB10 threads carry this fix already (amarrmb fork does too).
2. loaders.get_moshi_lm initializes the LM on the meta device (PR fix: reduce memory need during model init #18
   init-OOM fix). Any checkpoint key that is neither present nor covered by
   the dep_q 8->16 backfill patches survives
   load_state_dict(strict=False, assign=True) as a meta tensor and
   crashes the final .to() with "Cannot copy out of meta tensor". This
   fires for base-Moshi dep_q=8 checkpoints (depformer_emb.7.weight is
   not in the checkpoint and not covered by the 8..15 backfill).

Diff summary (isolated)

• moshi/pyproject.toml: torch >= 2.2.0, < 2.5 -> torch >= 2.2.0 (1 line).
• moshi/moshi/models/loaders.py: after the dep_q backfill, zero-init any
  model key still absent from the checkpoint, with a warning (13 lines).
  No behavior change for personaplex-7b-v1 checkpoints (all keys present).

Evidence

Verified on the env below: stock repo fails to pip-install (pin) and, with
the pin fixed, fails to load kyutai/moshiko-pytorch-bf16 (meta crash);
with this PR both work, and personaplex-7b-v1 loads byte-identically
(loader path unchanged when no keys are missing).

Pinned environment

• Lenovo ThinkStation PGX, NVIDIA GB10 (sm_121, 48 SMs, 25MB L2), aarch64,
  128GB unified LPDDR5X (~273 GB/s), 20-core Grace (Cortex-X925/A725)
• Ubuntu 24.04, driver 595.71.05, CUDA 13.0/13.2, python 3.12.3
• torch 2.13.0+cu130, triton 3.7.1 (bundled ptxas-blackwell replaced by
  system /usr/local/cuda/bin/ptxas 13.0.88 via symlink)

Tolerance-ladder tier

Tier 0 (build/load correctness; no numerics).

Known unknowns

sm_121, n=1 machine. The relaxed pin is unbounded above; maintainers may
prefer < 3.0. Not tested on sm_80/sm_90 (no hardware); the loader change
is inert when checkpoints are complete.

Attribution

• torch-pin relax: first shipped in @amarrmb's fork (commit 94cbbbd,
  github.com/amarrmb/personaplex); the pin conflict was reported in the
  GB10 discussion thread (PersonaPlex produces choppy/unusable audio on DGX Spark (GB10) #3, @acatovic). This PR
  carries the same one-line change with a Co-authored-by trailer.
• meta-tensor zero-init fix: authored here (jethac), interacting with
  upstream PR fix: reduce memory need during model init #18's meta-device init.
• @amarrmb: please flag any attribution adjustment you'd like — happy to
  amend.

Also in this branch

This branch (pr-build-fixes) now also carries @amarrmb's
encode_from_sphn dtype-cast bugfix, cherry-picked with authorship
preserved (their add7726) - it is a correctness fix independent of the
performance series. The performance work is consolidated into a single
second PR (pr-gb10-realtime, see 02-gb10-realtime.md), which is based
on this branch.

Upstream overlap (pre-flight scan, 2026-07-24)

• PR fix: widen dependency version ranges to resolve install failures #78 (@SuperMarioYL, Apr 8, open/unreviewed) widens ALL dependency
  version ranges in pyproject.toml and overlaps our torch-pin change.
  Credit to them for flagging the install failures broadly. Ours is
  deliberately narrower: a torch-only relax, since blanket-unpinning every
  dependency risks breakage from untested majors. We would happily rebase
  this PR on fix: widen dependency version ranges to resolve install failures #78 if the maintainer prefers the broader approach.
• PR fix: remove torch upper bound to support Python 3.13 #63 (@haosenwang1018, Feb 23) removes the same torch upper bound
  (motivated by Python 3.13 support) — same one-line intent as our pin
  relax; credit to them as an earlier report of the constraint. Either PR
  merging first makes that hunk of ours a no-op rebase.
• File-level only: Add macOS Nix and MLX runtime support #100/Fix client-side bugs, dependency pins, and misc issues #76/Add Smart UI Auto-Detection, Dynamic Voice System, and .env Support #37/Docker File and Pyproject.toml updates #32/Add real-time stdio streaming runtime with framed PCM protocol #61 also touch pyproject.toml for
  unrelated reasons (macOS/Nix, client fixes, UI, stdio runtime, Docker).

🤖 Generated with Claude Code