escha 1.2.0: docs + changelog for tensor parallelism

README.md

@@ -121,6 +121,27 @@ an answer is always produced: see
 
 ## Changelog
 
+**1.2.0** (2026-08-21) — **tensor parallelism (`--tp-size N`) now works.** The escha
+parameter class pins its own weight loader, which meant sglang's TP slicing never ran and
+every rank kept the whole checkpoint (rank 0 died with `weight must have shape
+(dim, width)`). It now slices per rank, including the fused-on-disk GDN `in_proj_qkv`,
+which is split into its three sub-projections first.
+
+> **Single-GPU users are unaffected.** Every new code path is gated on
+> `world_size > 1`; at `--tp-size 1` the loader is byte-for-byte what 1.1.1 did. Verified
+> as an identical shard layout and byte-identical greedy output.
+>
+> **TP > 1 is new and lightly tested — treat it as experimental.** It was contributed and
+> validated by [@ginerJuanUdesa](https://github.com/ginerJuanUdesa/escha-tp-fix-qwen3dense)
+> on 2× RTX 3090 (symmetric 6.02 GB/rank, coherent greedy output). **We have one GPU and
+> could not reproduce it**, and no numerical equivalence check against `--tp-size 1` has
+> been run yet. If you use it for evaluation, sanity-check a benchmark against the
+> single-GPU numbers first. Note that a multi-rank all-reduce reorders float accumulation,
+> so TP > 1 output is not expected to match TP = 1 bit-for-bit even when correct.
+>
+> On Ampere/Ada/Hopper you can add `DETERMINISTIC=1` to remove that reduction-order
+> variance if you want a stricter comparison.
+
 **1.1.1** (2026-08-21) — `process_weights_after_loading` now takes the rank's device
 instead of a hardcoded `cuda:0`. The hardcode put every 2-bit buffer on `cuda:0` while
 the input tensor sat on the server's actual device, so **any run not on device 0 —
@@ -130,11 +151,6 @@ traceback that pointed at the kernel rather than at the cause. Bit-identical whe
 `cuda:0` was already correct, which is every configuration `serve.sh` ships. Reported
 with a diagnosis and a fix by [@ginerJuanUdesa](https://github.com/ginerJuanUdesa/escha-tp-fix-qwen3dense).
 
-> Note: this fixes the device placement only. **`--tp-size > 1` is still not supported** —
-> the weight loader does not shard escha tensors per rank, so every rank holds the full
-> checkpoint. A community patch for that exists at the link above and is under review; it
-> is not in this wheel yet.
-
 **1.1.0** (2026-08-20) — first wheel with the dense (`escha`) serving path; the 1.0.x
 wheels registered `eschamoe` only, so a dense checkpoint failed at registry lookup.
 

THIRD_PARTY_LICENSES/THIRD_PARTY_NOTICES.md

@@ -1,4 +1,4 @@
-# Third-Party Notices — Escha Runtime (`escha` 1.1.1)
+# Third-Party Notices — Escha Runtime (`escha` 1.2.0)
 
 This repository distributes the **runtime** only (the `escha` wheel: quantization/serving
 kernels + a modified SGLang fork). SGLang is the only engine in this repository. The quantized **model weights** are distributed
@@ -9,7 +9,7 @@ All bundled third-party code is under **permissive licenses (Apache-2.0, MIT, BS
 an audit found **no copyleft (GPL / LGPL / AGPL / MPL / SSPL) code**. The obligations below are
 satisfied by the files in this `THIRD_PARTY_LICENSES/` directory together with the copyright
 notices retained verbatim in the redistributed source (inside the wheel). The same texts are also
-embedded in the wheel metadata (`escha-1.1.1+qwen3dense.dist-info/licenses/LICENSE`).
+embedded in the wheel metadata (`escha-1.2.0+qwen3dense.dist-info/licenses/LICENSE`).
 
 The `escha` runtime itself is released under the Apache License, Version 2.0 — see the top-level
 `LICENSE` file of this repository.

sglang/INSTALL.md

@@ -287,7 +287,7 @@ inject into the system prompt — at some cost in reasoning-dependent accuracy.
 - **`ValueError: Invalid quantization method ...`.** The wheel's registry ships exactly two
   methods, `escha` (dense) and `eschamoe` (mixture-of-experts). Check your model's
   `quantize_config.json` / `config.json` names one of them, and that you installed *this* wheel
-  (`python -c "import escha; print(escha.__version__)"` should print `1.1.1+qwen3dense`) — the
+  (`python -c "import escha; print(escha.__version__)"` should print `1.2.0+qwen3dense`) — the
   1.0.x wheels registered `eschamoe` only and cannot load a dense model at all.
 
 - **`ModuleNotFoundError: pybase64` / `msgspec` / `xgrammar` / `flashinfer` at startup.** The wheel