Add QNN (Qualcomm Hexagon HTP) execution-provider profile (#370)

## What
Registers a `qnn` execution-provider profile so the ORT-GenAI builder
can target the Qualcomm Hexagon NPU (HTP backend, via the
onnxruntime-qnn QAIRT plugin):
- `_execution_providers.py`: new `EpCapabilities(name="qnn", ...)`.
- `integrations/ort_genai/ep_config.py`: map `qnn` → `QNN` for ORT-GenAI
provider dispatch.

## Why
Bringing up gemma-4-12B (`gemma4_unified`) on Qualcomm Copilot+ devices
(Snapdragon X Elite / X2 Elite). The HTP runs a statically-shaped,
QDQ-quantized graph compiled to a QNN context binary and has no kernels
for ORT contrib fused ops, so the build must emit standard ONNX
primitives the QNN op builders can lower.

## How
The profile **reuses the existing rewrite infrastructure** rather than
adding new passes:
- `supports_fused_rope=False` → `SeparateRoPE` / `UnpackQKV`
- `supports_skip_layer_norm=False` → skip-norm inlined to standard ops
- `supports_packed_multi_head_attention=False` → PackedMHA expanded
- `gqa_dtypes=frozenset()` → no GroupQueryAttention (the gemma-4
multimodal decoder forgoes GQA anyway — bidirectional-vision overlay);
standard Attention is emitted and static-shaped downstream
- `supports_past_present_share_buffer=False` (standard-Attention KV
concat can't alias a shared buffer)
- `provider_options` = onnxruntime-qnn HTP launch defaults; `soc_model`
+ EP-context binary path are device-specific, set at build/validation
time.

## Validation
Building gemma-4-12B with `execution_provider="qnn"` emits pure ai.onnx
(zero `com.microsoft` contrib ops).

## Not in this PR (for discussion)
The HTP-specific lowering we currently do downstream — 4D q/k-norm
RMSNorm reshape, RoPE→`com.microsoft::RotaryEmbedding` op-swap,
fold-shapes / drop-identity / decoder-split / mask-hoist — is a
candidate follow-up PR. Some of it is QNN context-binary **deployment**
rather than model build; happy to bring the in-scope pieces in-tree as
`rewrite_rules`/`_passes` + tests if the team wants them here.

Heads-up: this sets GQA **off** for `qnn`; flagging against
`justinchu/gemma4-12b-text-gqa` in case the attention path needs to
reconcile.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>

Author: shreyshah-microsoft

src/mobius/_execution_providers.py

@@ -267,6 +267,28 @@ def _register_builtins() -> None:
             enable_graph_capture=True,
             supports_past_present_share_buffer=True,
         ),
+        # Qualcomm Hexagon NPU via the QNN EP (onnxruntime-qnn QAIRT plugin),
+        # HTP backend. The HTP runs a static-shaped, QDQ-quantized QNN context
+        # binary with no kernels for ORT contrib fused ops, so everything is
+        # decomposed to standard ONNX. The gemma4 multimodal decoder forgoes GQA
+        # (bidirectional-vision overlay), so standard Attention is emitted and
+        # static-shaped downstream. provider_options are the HTP launch defaults;
+        # soc_model and the EP-context binary path are set per-device at build time.
+        EpCapabilities(
+            name="qnn",
+            gqa_dtypes=frozenset(),  # no GroupQueryAttention (no QNN GQA builder)
+            qkv_pack_dtypes=frozenset(),  # no PackQKV
+            supports_fused_rope=False,  # SeparateRoPE + UnpackQKV
+            supports_skip_layer_norm=False,  # inline Skip[Simplified]LayerNorm
+            supports_packed_multi_head_attention=False,  # inline PackedMHA
+            provider_options={
+                "backend_path": "QnnHtp.dll",
+                "htp_performance_mode": "burst",
+                "htp_graph_finalization_optimization_mode": "3",
+                "enable_htp_shared_memory_allocator": "1",
+            },
+            supports_past_present_share_buffer=False,  # standard-Attention KV concat
+        ),
         # onnx-standard: ONNX-only runtime — emits zero custom-domain ops.
         # All com.microsoft ops (SkipLayerNorm, PackedMHA) are expanded via
         # InlinePass to their standard-ONNX function bodies. No GQA or QKV

src/mobius/integrations/ort_genai/ep_config.py

@@ -26,6 +26,7 @@
     "dml": "dml",
     "webgpu": "webgpu",
     "trt-rtx": "NvTensorRtRtx",
+    "qnn": "QNN",
 }