Add VarlenNGramHashMapping for packed/varlen batching support in DeepSeek Engram

[Copilot]

Description

NGramHashMapping operates on a dense (batch_size, sequence_length) layout and mixes n-gram hash inputs across positions within a row. Packing multiple sequences into one flat buffer (as required by ragged/packed batching engines like OgaEngine) and reshaping to (1, num_tokens) would silently leak hash-mix inputs across sequence boundaries at every packed request boundary.

This adds a new com.microsoft op, VarlenNGramHashMapping, modeled on VarlenCausalConvWithState, which accepts a cumulative_sequence_length (shape (batch_size + 1), device-resident int32) offsets tensor and clamps the n-gram window — and past_ids/present_ids history — at each packed request's own boundary instead of the flat buffer's.

• Schema: New op in bert_defs.cc (registered in ms_opset.h) taking token-major packed input_ids ((total_tokens)) plus cumulative_sequence_length; past_ids/present_ids are now indexed per-request instead of per fixed-stride batch row.
• CPU kernel: Parallelizes over requests via ThreadPool::TryParallelFor; validates offsets host-side.
• CUDA kernel: Block-per-request design (mirrors VarlenCausalConvKernel); each block validates its own offsets before touching memory.
• WebGPU kernel: Workgroup-per-request dispatch for hash ids, with a separate program for present_ids.
• Prerequisite port: NGramHashMapping/EngramGate (from unmerged PR Add DeepSeek Engram contrib ops (EngramGate, NGramHashMapping) #32268) were not yet on this branch; selectively ported only the NGramHashMapping-related CPU/CUDA/WebGPU kernels, engram_helper files, and tests needed here (EngramGate and unrelated CausalConvWithState dilation/channels_last changes are out of scope).
• Docs: Hand-updated docs/ContribOperators.md and docs/OperatorKernels.md to match gen_doc output.
• Tests: Added coverage in engram_ops_test.cc for packed-vs-per-sequence equivalence, sequence-boundary no-leakage, packed chunked/decode equivalence across concurrent requests, negative ids, and invalid vocab_sizes.

# Equivalent results, but VarlenNGramHashMapping never reads across the sequence boundary
NGramHashMapping(seq_a) + NGramHashMapping(seq_b)
  == VarlenNGramHashMapping(concat(seq_a, seq_b), cumulative_sequence_length=[0, len(seq_a), len(seq_a)+len(seq_b)])

Motivation and Context

NGramHashMapping's dense (batch_size, sequence_length) layout is incompatible with OgaEngine's packed/ragged batching model (continuous batching / paged attention), where tokens from multiple in-flight requests are concatenated into a single flat, unpadded buffer. Without this change, models using DeepSeek Engram would silently corrupt n-gram hash ids at every sequence boundary whenever more than one request is batched together.

[azure-pipelines]

Azure Pipelines:
There may be pipelines that require an authorized user to comment /azp run to run.

[Tianlei Wu (tianleiwu)]

The packed-boundary design and the follow-up offset-safety fixes look solid, and the multi-EP tests are strong. I found one blocking WebGPU bug when the optional present_ids output is omitted, plus one shape-inference validation gap for the same one-output form. Details are inline.

[Copilot]

🟡 Changes recommended

Shape validation incorrectly depends on an optional output, and long prefills are serialized across all three implementations.

Once you've addressed the issues Copilot identified, you can request another Copilot review.

Review details

• Files reviewed: 18/18 changed files
• Comments generated: 4
• Review effort level: Balanced