Upload README.md with huggingface_hub

README.md

@@ -0,0 +1,62 @@
+---
+license: apache-2.0
+tags: [qhexrt, hexagon, npu, qnn, llm, gated-delta-net]
+base_model: Qwen/Qwen3.5-4B
+pipeline_tag: text-generation
+---
+
+# Qwen3.5-4B — Hexagon NPU (QHexRT) — v81
+
+A prebuilt **QHexRT** bundle of **Qwen/Qwen3.5-4B** (the **Qwen3-Next "GatedDeltaNet" hybrid** lineage)
+running on the **Qualcomm Hexagon v81 NPU** (SM8850). text→text.
+
+The text decoder is a **hybrid**: 3 of every 4 layers are a **recurrent gated delta-rule linear attention**
+(`Qwen3_5GatedDeltaNet`: a short causal-conv FIFO + an SSM state + gated RMSNorm), and every 4th layer is
+**gated softmax attention** (q-proj output gate, per-head q/k RMSNorm, **partial-rotary** M-RoPE θ=1e7). The
+runtime drives it with the `qwen3_5_generate` host-op (decode-over-prompt; conv + SSM + KV state carried
+host-side), the lm-head as a separate graph.
+
+**Multi-context decode.** The 32-layer W8 decode (3.6 GB) exceeds the v81 cDSP **per-context** memory ceiling
+(`contextCreateFromBinary → 0x3ea MEM_ALLOC`). So the decode is split into **2 context shards** by layer range
+(0–15 / 16–31, ~1.81 GB each); the host-op runs them in sequence, threading the residual stream shard→shard
+(only the last shard applies the final norm). Each shard updates only its own layers' conv/SSM/KV state. This
+is numerically identical to a single decode graph — parity stays greedy-exact.
+
+## What's here (`v81/` — the flat artifacts root)
+| file | role | ~size |
+|---|---|---|
+| `qwen3.5-4b-1024.json` | QHexRT manifest (the declarative run plan) | 2 KB |
+| `qwen354b_decode_s0_w8.bin` | decode shard 0 (layers 0–15, W8) | 1.81 GB |
+| `qwen354b_decode_s1_w8.bin` | decode shard 1 (layers 16–31, W8) | 1.81 GB |
+| `qwen354b_lmhead_f16.bin` | lm-head context binary (tied embed, fp16) | 1.27 GB |
+| `qwen354b_embed_f16.bin` | embedding table (host lookup, fp16) | 1.27 GB |
+| `tokenizer.json` | Qwen2-style BPE tokenizer | 20 MB |
+
+Arch-pinned: a v81 binary will not load on another Hexagon arch. The QNN runtime libs come from the QAIRT SDK,
+not this repo.
+
+## Run
+```bash
+hf download runanywhere/qwen3_5_4b_HNPU --local-dir q35_4b
+adb push q35_4b/v81 /data/local/tmp/wq/qwen35-4b
+# also stage the QAIRT runtime libs + a qhx_generate built with the qwen3_5_generate host-op (see caveats)
+adb shell "cd /data/local/tmp/wq && LD_LIBRARY_PATH=. \
+  ./qhx_generate qwen35-4b/qwen3.5-4b-1024.json libQnnHtp.so libQnnSystem.so qwen35-4b 64 'The capital of France is'"
+```
+
+## Measured (device: SM8850 / v81, QAIRT 2.47)
+- **Decode: ~5.5 tok/s** (W8 weight-only, 2 context shards).
+- **Parity: greedy-EXACT** — on-device tokens match the HF `Qwen/Qwen3.5-4B` greedy continuation **64/64**
+  (prompt "The capital of France is").
+
+## Caveats
+- **W8 weight-only** decode (the ~7 GB fp16 decode exceeds device RAM; W8 ~3.6 GB) + **fp16** lm-head. Greedy
+  parity is exact under W8 (64/64) for this model.
+- **Multi-context**: the decode is 2 shards (`decode_s0` + `decode_s1`) plus the lm-head — 3 context binaries
+  loaded together. A single decode context this size fails on v81 (per-context alloc ceiling, `0x3ea`).
+- The W8 decode `.so` (per shard ~1.8 GB) is built with a **large-code-model** path (weight blob → `.lrodata`
+  + `-mcmodel=large`); the lm-head ONNX is built directly (MatMul + external fp16 weight) to dodge the 2 GB
+  protobuf limit. Both are baked into the published bins.
+- **Text-LLM path only** — the vision tower + MTP head are not exported.
+- Needs a `qhx_generate` that includes the **`qwen3_5_generate`** host-op (QHexRT branch `smonga/qwen_fam`).
+- Built by the in-repo `forge` pipeline (oracle-gated export 10/10 → QAIRT-2.47 O3 compile → device greedy gate).