issue/1042 - fix address misaligned error of rearrange

[pengcheng888]

`import infinicore

aa = infinicore.empty((4, 3), dtype=infinicore.int64, device=infinicore.device("cuda"))
bb = infinicore.empty((4, 2), dtype=infinicore.int64, device=infinicore.device("cuda"))
cc = aa.narrow(1, 0, 2)

print(aa.shape, aa.stride())
print(bb.shape, bb.stride())
print(cc.shape, cc.stride())

bb.copy_(cc)

infinicore.sync_stream()`

上述样例代码会报错。

测试结果：

修改前

修改后

算子单测

服务推理结果
cuda_graph的推理不再出错

[pengcheng888]

服务启动代码
`export CUDA_VISIBLE_DEVICES=1

xmake build _infinilm && xmake install _infinilm

python python/infinilm/server/inference_server.py
--nvidia
--model_path=/data-aisoft/mechdancer/models/9g_8b_thinking_llama/
--max_batch_size 6
--num_blocks 8192
--tp 1
--temperature 1.0
--block_size 16
--top_p 0.8
--top_k 1
--port 8102
--max_tokens 256
--enable-graph
--cache_type paged
`

客户端启动代码
`export CUDA_VISIBLE_DEVICES=1
unset http_proxy https_proxy all_proxy ALL_PROXY

vllm bench serve --backend openai-chat --model 9g_8b_thinking
--endpoint /chat/completions
--tokenizer /data-aisoft/zenghua/models/9g_8b_thinking_llama
--max-concurrency 16
--request_rate 10.0
--port 8102
--seed $(date +%s)
--ready-check-timeout-sec 3000
--num-prompts 64
--random-input-len 16
--extra-body '{"max_tokens": 64}'`

[Copilot]

Copilot encountered an error and was unable to review this pull request. You can try again by re-requesting a review.

[Copilot]

Pull request overview

Copilot reviewed 1 out of 1 changed files in this pull request and generated 2 comments.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

The is_contiguous function has an infinite loop bug. The loop variable i is declared as size_t (an unsigned type), and the condition i >= 0 is always true. When i reaches 0 and is decremented with --i, it wraps around to SIZE_MAX due to unsigned underflow, causing an infinite loop (or out-of-bounds memory access). The loop counter should use a signed type (e.g., ptrdiff_t) to allow the termination condition to work correctly.

Suggested change

	for (size_t i = ndim - 1; i >= 0; --i) {
	if (ndim == 0) {
	return true;
	}
	for (ptrdiff_t i = static_cast<ptrdiff_t>(ndim) - 1; i >= 0; --i) {

[Copilot]

The is_contiguous function is called at line 68 with ndim that has already been reassigned to dims.size() (at line 64), which is the number of dimensions after filtering out size-1 dimensions. However, the function receives the original shape and src_strides_/dst_strides_ pointer arrays (from the caller), which still have the original, unfiltered length. This mismatch means is_contiguous iterates over only the last dims.size() entries of the original arrays instead of all entries, potentially reporting a false contiguous result. The contiguity check should use the original ndim (before line 64 reassigns it) or iterate over the dims vector instead.

[pengcheng888]

@copilot code review[agent]

[Copilot]

Pull request overview

Copilot reviewed 1 out of 1 changed files in this pull request and generated 2 comments.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

The alignment check only validates the first ([0]-th) element of dst_strides and src_strides. For correct alignment, all outer strides must be multiples of the chosen new_unit, not just the first one. When ndim > 1, intermediate strides at indices 1 through ndim-1 are not checked.

For example, if ndim == 2 and dst_strides == [40, 12] with candidate=8, then 40 & 7 == 0 passes the check but 12 & 7 == 4 != 0 means stride index 1 is not 8-byte aligned, leading to misaligned access.

The loop over candidates should check alignment for ALL strides (dst_strides()[0..ndim-1] and src_strides()[0..ndim-1]), not just index 0.

[pengcheng888]

有必要检查每一个维度的stride么

[Copilot]

When ndim == 0 (all dimensions were merged into unit during RearrangeMeta::create()), _meta has only 2 elements (indices 0 and 1). In that case, _meta[2 + ndim] = _meta[2] and _meta[2 + ndim + ndim] = _meta[2] are out-of-bounds accesses, causing undefined behavior.

A guard should be added so that when ndim == 0, the alignment check on strides is skipped entirely (there are no strides to check). For example, the alignment checks should only be performed when ndim > 0, and when ndim == 0, only the current_unit % candidate == 0 condition needs to be satisfied.

[pengcheng888]

在这个函数中，ndim会永远大于0，不会存在ndim == 0的情况

[wooway777]

resolves #1042