Emit bias-aware external-KV (TensorScatter + nonpad_kv_seqlen) attention for float-bias decoders

[titaiwangms]

Problem

Decoders that require a float additive attention bias — Gemma-4's bidirectional vision-block overlay, and sliding-window / custom-mask models — cannot use com.microsoft.GroupQueryAttention (it accepts only causal/local masking, not an arbitrary attention_bias). mobius therefore emits them with the standard ONNX Attention op, is_causal=0, over an internal dynamic past/present KV cache. With a bias present this routes ORT to the internal-cache MEA path, which the kernel's own source note calls ~15-30% slower than contrib GQA's in-place decode. Today these models give up the decode fast-path entirely.

ORT supports a faster alternative on the SAME standard Attention op: the opset-24 external KV cache (nonpad_kv_seqlen + TensorScatter). On the MEA path it combines additive bias + GQA + external cache and reaches near-contrib-GQA performance — letting a single standard-Attention graph carry the bias during prefill AND hit a near-GQA decode (Flash is precluded by any bias, so MEA-external ≈ GQA is the ceiling).

Mobius-side deliverable (this issue — contract-independent, ORT-testable now)

mobius already emits the maskless external-KV static-cache graph (sibling work #345) and already builds the causal+sliding+Gemma-4 block-overlay bias (create_attention_bias). This issue wires the two together:

• Thread an optional float bias into the shared static-cache Attention path (_apply_attention), pairing bias-present with is_causal=0 (maskless is_causal=1 stays the default).
• Add an external-cache bias builder producing a (B,1,S_q,max_seq) additive bias keyed on absolute query positions with KV validity slot < nonpad_kv_seqlen, reusing the proven causal/sliding/block-overlay/padding logic.
• Flag-gated (MOBIUS_STATIC_CACHE_BIAS, default off).
• An ORT manual-feed parity test (CPU MEA, no genai, no Flash) vs a dense-attention reference for a prefill chunk and a decode step.
  The bias geometry is a pure function of model semantics + cache layout — independent of how the runtime drives it — so this slice is safe to land before the genai consumer exists.

Acceptance criteria

• Shared static-cache Attention path honours an optional float bias with is_causal=0; maskless default unchanged.
• create_static_cache_attention_bias emits the (B,1,S_q,max_seq), slot<nonpad bias (causal + sliding + block-overlay + padding).
• ORT-level parity test passes on CPU (MEA) for prefill + decode vs a dense reference; matches HF Gemma-4 single-layer attention.
• Flag-gated, default off; no shipped model's emission changes when off.

Explicitly out of scope (separate follow-ups)

Re-emitting Gemma-4's own decoder task/model on the static path, genai_config.json wiring, single-graph vs prefill/decode-split packaging, and e2e generation/decode profiling.

Relationships

• Part of Emit opset-24 external KV cache (nonpad_kv_seqlen) for bias decoders to recover near-GQA decode (Gemma-4, sliding-window) #349 (Gemma-4 end-to-end near-GQA decode consumer).
• Sibling to Emit maskless is_causal=1 + nonpad_kv_seqlen static-cache graph once upstream lands #345 (maskless is_causal=1 static-cache graph) — shares the TensorScatter + nonpad_kv_seqlen primitive; this issue adds the bias variant.
• genai-side driving tracked separately: Data-dependent bidirectional mask (Gemma-4 vision-block) forces standard Attention over GQA — split prefill/decode decoder graphs? microsoft/onnxruntime-genai#2204 (Path A single external-KV graph vs Path B prefill/decode split). This issue is contract-independent and unblocks regardless of which path genai picks.