refactor(ops): name Infinilm replacements

[voltjia]

Summary

• Replace generic deprecation diagnostics on seven Infinilm compatibility operators with their concrete canonical replacements.
• Keep the legacy classes, signatures, providers, and tests unchanged.
• Leave the four interfaces without a complete single-operator replacement on the generic migration diagnostic.

Motivation

The canonical destinations for convolution, paged cache writes, decode/prefill attention, copies, batched top-k/top-p sampling, and zero filling are now present on master. Their deprecated compatibility APIs should name those destinations directly so compiler diagnostics provide actionable migration guidance.

No issue is linked. PR #893 was closed separately after the post-#911 audit because it targeted the retired vLLM v0.6.3 paged_attention_v1 wrapper and duplicated the maintained FlashAttention replacement path.

Type of Change

• feat - new feature / new operator / new platform
• fix - bug fix
• perf - performance improvement (no behavioral change)
• refactor - code restructuring without behavior change
• test - adding or fixing tests only
• docs - documentation only
• build / ci - build system or CI configuration
• chore - tooling, formatting, or other non-code changes
• Breaking change (requires a ! in the Conventional Commits prefix or a BREAKING CHANGE: footer)

Platforms Affected

• CPU (WITH_CPU)
• NVIDIA (WITH_NVIDIA)
• Iluvatar (WITH_ILUVATAR)
• MetaX (WITH_METAX)
• Cambricon (WITH_CAMBRICON)
• Moore (WITH_MOORE)
• Ascend (WITH_ASCEND)
• PyTorch C++ bindings (WITH_TORCH)
• Build system / CMake / CI
• Python bindings / user-facing API

Only compile-time deprecation text changes. Runtime behavior is unchanged on every platform.

Smoke Test Result

Remote environment: ssh nvidia, image accelerator-dev/nvidia:latest, with the existing InfiniRT prefix and /tmp CUTLASS source.

$ CMAKE_BUILD_PARALLEL_LEVEL=8 python3 -m pip install . \
--no-build-isolation --no-deps \
-C cmake.define.INFINI_RT_ROOT=/opt/infinirt \
-C cmake.define.AUTO_DETECT_DEVICES=OFF \
-C cmake.define.AUTO_DETECT_BACKENDS=OFF \
-C cmake.define.WITH_CPU=ON \
-C cmake.define.WITH_NVIDIA=ON \
-C cmake.define.WITH_LINKED=OFF \
-C cmake.define.WITH_TORCH=OFF \
-C cmake.define.FETCHCONTENT_SOURCE_DIR_CUTLASS=/opt/cutlass \
-C cmake.define.INFINI_OPS_OPS=conv_infinilm,paged_attention_infinilm,paged_attention_prefill_infinilm,paged_caching_infinilm,rearrange_infinilm,top_k_top_p_sample_infinilm,zeros_infinilm
Successfully built InfiniOps
$ python3 -m pytest \
tests/test_conv_infinilm.py \
tests/test_paged_attention_infinilm.py \
tests/test_paged_attention_prefill_infinilm.py \
tests/test_paged_caching_infinilm.py \
tests/test_rearrange_infinilm.py \
tests/test_top_k_top_p_sample_infinilm.py \
tests/test_zeros_infinilm.py \
--devices cpu cuda -q
236 passed, 192 skipped in 16.68s

The skips are the existing unsupported device/operator combinations in the legacy test matrix.

Test Results on Supported Platforms

Platform	Affected	Build / Smoke Result	Full Result / Notes
NVIDIA	Yes	focused build and tests passed	Included in 236 passed; no runtime change
Iluvatar	Yes	strict wrapper generation passed	Header-only diagnostic change; hardware not run
MetaX	Yes	strict wrapper generation passed	Header-only diagnostic change; hardware not run
Cambricon	No	N/A - not affected	N/A - no provider for these compatibility APIs
Moore	Yes	strict wrapper generation passed	Header-only diagnostic change; hardware not run
Ascend	Yes	strict wrapper generation passed	Header-only diagnostic change; hardware not run

Additional checks:

clang-format 21.1.8 --dry-run --Werror <seven changed headers>
passed
python3 scripts/generate_wrappers.py \
--devices nvidia iluvatar metax moore \
--ops conv_infinilm paged_attention_infinilm \
paged_attention_prefill_infinilm paged_caching_infinilm \
rearrange_infinilm zeros_infinilm \
--strict-ops
passed
python3 scripts/generate_wrappers.py \
--devices cpu ascend \
--ops top_k_top_p_sample_infinilm \
--strict-ops
passed
python3 -m pytest tests/test_generate_wrappers.py -q
25 passed in 1.08s

Benchmark / Performance Impact

N/A. No executable implementation changes.

Notes for Reviewers

Replacement alignment

Deprecated API	Canonical replacement	Open-source alignment evidence	Migration note
ConvInfinilm	Convolution	PyTorch torch.convolution and fixed ATen schema	Existing adapter supplies transposed=false, zero output_padding, and reorders stride/padding.
PagedCachingInfinilm	ReshapeAndCacheFlash	vLLM reshape_and_cache_flash	Consumers must adopt the canonical FlashAttention cache layout.
PagedAttentionInfinilm	FlashAttnWithKvcache	FlashAttention flash_attn_with_kvcache	Consumers use the matching canonical cache layout; the NVIDIA provider is linked from the installed DSO.
PagedAttentionPrefillInfinilm	FlashAttnVarlenFunc	FlashAttention flash_attn_varlen_func	Paged KV, block tables, and ALiBi are supported by the linked provider merged in #911.
RearrangeInfinilm	Copy	PyTorch Tensor.copy_	The old contract is the non_blocking=false subset.
TopKTopPSampleInfinilm	TopKTopPSamplingFromLogits	FlashInfer top_k_top_p_sampling_from_logits	Callers materialize per-batch top_k / top_p tensors and pass the canonical explicit attributes.
ZerosInfinilm	Fill with value 0	PyTorch Tensor.fill_	InfiniOps retains the caller-provided trailing output convention.

CausalSoftmaxInfinilm, ScaledSoftmaxInfinilm, KvCachingInfinilm, and RandomSampleInfinilm retain the generic diagnostic because no single stable public operator completely represents their existing contracts. This PR does not invent an approximate replacement or add an overload.

[voltjia]

Closing per scope prioritization: this PR only improves deprecation diagnostics and is not required for the functional InfiniLM migration. Continuing with the remaining executable migration work first.