Summary

This PR extends the NPU FlashComm1 sequence-parallel path introduced in
#1890 to DeepSeek V4 /
DSV4-Flash.

During prefill, hidden states are sharded across the TP group and kept in the
sharded layout where supported. The full-sequence layout is restored before
DeepSeek Sparse Attention and routed-MoE computation, then converted back to
the sharded layout through ReduceScatter or explicit re-sharding. The final
hidden states are gathered before the model head and MTP state handling.

FlashComm1 remains disabled by default:

--enable_flashcomm1=false

Enable it with:

--enable_flashcomm1=true \
--flashcomm1_min_prefill_tokens=1000

MMRS fusion remains an independent opt-in inherited from #1890:

--enable_mmrs_fusion=true \
--mmrs_comm_mode=aiv

The supplied performance record identifies only the FlashComm1 off/on toggle;
it does not identify the MMRS setting or communication mode. Therefore, the
results below make no separate MMRS performance claim.

Main Changes

• Integrate FlashComm1Context into the DeepSeek V4 NPU prefill path.
• Shard model hidden states before the decoder layers and gather them before
  model-output and MTP state handling.
• Restore the full-sequence layout before DeepSeek Sparse Attention and
  routed-MoE computation.
• Use the FlashComm1 row-parallel reduce mode for the DeepSeek Sparse Attention
  output projection.
• Re-shard routed-MoE output before returning to the sequence-parallel decoder
  layout.
• Exclude the shared-expert DenseMLP from the generic FlashComm1
  gather/reduce-scatter path, avoiding an incorrect gather inside the
  shared-expert branch.
• Remove the Qwen3.5-specific dp=1 and tp>=8 activation restrictions so the
  tested dp4 tp4 topology can use FlashComm1.
• Change the global default prefill-token threshold from 8192 to 1000. This
  policy change also applies to the existing Qwen3.5 FlashComm1 path.
• Add unit coverage for eligibility, context scope, sequence sharding, tail
  padding, and reduce-mode selection.

Default Activation Policy

FlashComm1 becomes active only when all applicable runtime conditions are met:

• NPU backend.
• --enable_flashcomm1=true.
• Prefill-only forward.
• Current prefill token count is at least
  --flashcomm1_min_prefill_tokens, which now defaults to 1000.
• cp_size=1; context parallelism remains unsupported.
• A TP process group is available. More than one TP rank is required for
  actual sequence sharding.
• The current DeepSeek V4 forward is not an ACL Graph forward.
• The current DP rank is not empty.

There is no longer an explicit dp=1 or tp>=8 requirement. Decode remains
outside the FlashComm1 path.

ACL Graph was enabled in the benchmark deployment, but the implementation does
not run FlashComm1 inside an ACL Graph forward. FlashComm1 operates on the
eligible non-graph prefill path, while graph execution can remain enabled for
the deployment.

Validation

Unit coverage added

• FlashComm1 eligibility for dp4 tp4.
• Minimum-prefill-token and prefill-only gating.
• Rejection when context parallelism is enabled.
• Thread-local context publication and restoration.
• Divisible sequence sharding.
• Tail-rank-only padding for uneven token counts.
• Row padding and zero-fill behavior.
• ReduceScatter versus MMRS reduce-mode selection.

Checks performed

• pre-commit run --all-files: passed.
• Performance benchmark: completed with the supplied results below.

Performance

Test Configuration

Measured Results

At concurrency 32, total throughput per card increases from 604.57 to 646.43.
At concurrency 1, it increases from 50.43 to 50.76.