Add QNN (Qualcomm Hexagon HTP) execution-provider profile

Registers an EpCapabilities entry for the "qnn" execution provider so the
ORT-GenAI builder can target the Qualcomm Hexagon NPU (HTP backend, via the
onnxruntime-qnn QAIRT plugin), and maps "qnn" -> "QNN" in the ORT-GenAI
provider-name table.

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 profile
forces decomposition to the standard ONNX primitives the QNN op builders can
lower: no fused RoPE, no Skip[Simplified]LayerNorm, no PackedMHA, and no
GroupQueryAttention (empty gqa_dtypes) -- the gemma4 multimodal decoder
forgoes GQA anyway (bidirectional-vision overlay). past_present_share_buffer
is disabled since standard-Attention KV concat cannot alias a shared buffer;
fixed prefill-bucket / fixed-max-KV static shaping is applied as a separate
downstream pass before context-binary generation.

Validated by building gemma-4-12B (gemma4_unified) with
execution_provider="qnn": the graph emits pure ai.onnx (zero com.microsoft
contrib ops).

Author: shreyshah-microsoft

src/mobius/_execution_providers.py

@@ -267,6 +267,47 @@ def _register_builtins() -> None:
             enable_graph_capture=True,
             supports_past_present_share_buffer=True,
         ),
+        # Qualcomm Hexagon NPU via the QNN execution provider (the
+        # onnxruntime-qnn QAIRT plugin EP), targeting the HTP backend. The HTP
+        # runs a STATICALLY-shaped, QDQ-quantized graph compiled into a QNN
+        # context binary, and has no kernels for ORT contrib fused ops — so we
+        # decompose everything down to standard ONNX primitives the QNN op
+        # builders / HTP can lower:
+        #   supports_fused_rope=False        -> SeparateRoPE + UnpackQKV
+        #                                       (no com.microsoft RotaryEmbedding
+        #                                        fused into attention).
+        #   supports_skip_layer_norm=False   -> InlinePass expands Skip[Simplified]
+        #                                       LayerNormalization to standard ops.
+        #   supports_packed_multi_head_attention=False -> InlinePass expands
+        #                                       PackedMultiHeadAttention.
+        #   gqa_dtypes=frozenset()           -> no GroupQueryAttention op: QNN has
+        #                                       no GQA builder, and the gemma4
+        #                                       multimodal decoder forgoes GQA
+        #                                       anyway (bidirectional-vision
+        #                                       overlay). Standard Attention is
+        #                                       emitted and static-shaped downstream.
+        # KV cache therefore uses standard-Attention concat semantics (share-buffer
+        # aliasing requires GQA), so supports_past_present_share_buffer=False; the
+        # fixed prefill-bucket / fixed-max KV shaping is applied as a separate
+        # static-shape pass before QNN context-binary generation.
+        # provider_options are the onnxruntime-qnn HTP launch defaults; soc_model
+        # and the EP-context binary path are device-specific and are set at
+        # build/validation time on the Snapdragon target.
+        EpCapabilities(
+            name="qnn",
+            gqa_dtypes=frozenset(),
+            qkv_pack_dtypes=frozenset(),
+            supports_fused_rope=False,
+            supports_skip_layer_norm=False,
+            supports_packed_multi_head_attention=False,
+            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,
+        ),
         # 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 +26,9 @@
     "dml": "dml",
     "webgpu": "webgpu",
     "trt-rtx": "NvTensorRtRtx",
+    # ORT-GenAI normalizes "qnn" -> "QNN" (config.cpp NormalizeProviderName) and
+    # dispatches it to QNNExecutionProvider::AppendExecutionProvider.
+    "qnn": "QNN",
 }