Upload folder using huggingface_hub

.gitattributes

@@ -33,3 +33,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+v81/tokenizer.json filter=lfs diff=lfs merge=lfs -text

README.md

@@ -0,0 +1,52 @@
+---
+license: other
+license_name: nvidia-open-model-license
+base_model: nvidia/llama-3.2-nv-embedqa-1b-v2
+tags: [qhexrt, hexagon, npu, qualcomm, embedding, on-device, v81, snapdragon]
+library_name: qhexrt
+pipeline_tag: feature-extraction
+---
+
+# NV-EmbedQA-1B → Hexagon NPU (QHexRT, v81)
+
+On-device text-embedding bundle of **[nvidia/llama-3.2-nv-embedqa-1b-v2](https://huggingface.co/nvidia/llama-3.2-nv-embedqa-1b-v2)**
+for the **Qualcomm Hexagon v81 NPU** (SM8850 / Snapdragon 8 Elite Gen-class), runnable with the
+[QHexRT](https://github.com/RunanywhereAI) C++ runtime. **No Python in the hot path.**
+
+The model is a `LlamaBidirectionalModel` (rotary Llama-3.2-1B layers, **full/bidirectional** attention,
+hidden 2048, 16 layers, 32 query / 8 KV heads, head_dim 64) followed by **avg-pool over valid tokens + L2**.
+It produces a **2048-d** sentence embedding.
+
+## Device validation (v81, SM8850)
+| metric | value |
+|---|---|
+| **device-vs-HF cosine** (8-tok query) | **1.00000** |
+| **device-vs-HF cosine** (8-tok doc) | **1.00000** |
+| **device-vs-HF cosine** (33-tok input) | **1.00000** |
+| retrieval cos(q,d): device / HF | 0.83613 / 0.83598 |
+| encoder latency | ~36 ms |
+
+The on-device embedding is **bit-faithful to the HuggingFace reference** across input lengths (cosine ≥ 0.9999).
+
+## What's inside (`v81/`)
+- `nvembedqa_enc.bin` — the bidirectional Llama encoder as a v81 fp16 QNN context binary (graph
+  `nvembedqa_enc_f16`; host-computed RoPE fed as inputs, **no in-graph rotary Gather**; seq 64).
+- `nvembedqa_embed_f16.bin` — the token-embedding table (fp16).
+- `tokenizer.json`, `tokenizer_config.json` — the Llama-3 BPE tokenizer.
+- `nv-embedqa-1b.json` — the QHexRT manifest (host-op `llama_embed`: tokenize → embed → llama3-scaled RoPE
+  (factor 32) → bidirectional padding mask → encoder graph → avg-pool → L2).
+
+## Run
+```bash
+adb push v81 /data/local/tmp/wq/nvembed
+adb shell "cd /data/local/tmp/wq && LD_LIBRARY_PATH=. \
+  QHX_EMB_QUERY_PREFIX= QHX_EMB_DOC_PREFIX= \
+  ./qhx_embed nvembed/nv-embedqa-1b.json libQnnHtp.so libQnnSystem.so nvembed \
+  'What is the capital of France?' 'Paris is the capital of France.'"
+# -> QUERY/DOC 2048-d L2-normalized vectors + cosine(query, document)
+```
+Embeddings are computed over **raw text** (no task prefix) — set the prefix env vars to empty as shown.
+
+## Arch pinning
+A QNN context binary is **arch- and QAIRT-version-pinned**: this `v81` bundle (QAIRT 2.47, soc_model 87) loads
+only on a v81 device whose runtime libs are ≥ the compile QAIRT. It will **not** load on v79/v83.

v81/nv-embedqa-1b.json

@@ -0,0 +1,39 @@
+{
+  "schema_version": 1,
+  "_comment": "NV-EmbedQA-1B (LlamaBidirectional + avg-pool) on Hexagon v81. Host-rope bidirectional encoder graph (no in-graph Gather) + llama_embed host-op (embed + llama3 rope factor 32 + bidirectional mask + avg-pool + L2). Encoder S=64.",
+  "model": {
+    "name": "nv-embedqa-1b",
+    "family": "embedding",
+    "dsp_arch": "v81",
+    "tokenizer_pre": "llama3"
+  },
+  "params": {
+    "hidden": 2048,
+    "vocab": 128256,
+    "n_layers": 16,
+    "head_dim": 64,
+    "max_ctx": 64,
+    "eos_token_id": 128001
+  },
+  "artifacts": {
+    "contexts": {
+      "encoder": {
+        "bin": "nvembedqa_enc.bin"
+      }
+    },
+    "embed": "nvembedqa_embed_f16.bin",
+    "tokenizer": "tokenizer.json"
+  },
+  "plan": {
+    "steps": [
+      {
+        "host": "llama_embed",
+        "params": {
+          "encoder": "nvembedqa_enc_f16",
+          "rope_factor": 32.0,
+          "neg": -50000
+        }
+      }
+    ]
+  }
+}

v81/nvembedqa_embed_f16.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d0f26c68ed40afd43c2510ef455064d3b4738765b5ea39d2cf48648b5e245cf6
+size 525336576

v81/nvembedqa_enc.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:469c2414679694612dc48406800a4572cb9994fe3f48938a94904bb3ed7ec486
+size 1950478336

v81/tokenizer.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6b9e4e7fb171f92fd137b777cc2714bf87d11576700a1dcd7a399e7bbe39537b
+size 17209920

v81/tokenizer_config.json

@@ -0,0 +1,15 @@
+{
+  "backend": "tokenizers",
+  "bos_token": "<|begin_of_text|>",
+  "clean_up_tokenization_spaces": true,
+  "eos_token": "<|end_of_text|>",
+  "is_local": false,
+  "local_files_only": false,
+  "model_input_names": [
+    "input_ids",
+    "attention_mask"
+  ],
+  "model_max_length": 131072,
+  "pad_token": "<|end_of_text|>",
+  "tokenizer_class": "TokenizersBackend"
+}