[Optimization] [OP] [Models] dsk del prefill mask

[chang-wenbin]

Motivation

DeepSeek V3 模型性能优化，包括 rotary kernel 支持 >65535 token、merge 算子支持多 head_dim

💡 If this PR is a Cherry Pick, the PR title needs to follow the format by adding the [Cherry-Pick] label at the very beginning and appending the original PR ID at the end. For example, [Cherry-Pick][CI] Add check trigger and logic(#5191)

💡 如若此PR是Cherry Pick，PR标题需遵循格式，在最开始加上[Cherry-Pick]标签，以及最后面加上原PR ID，例如[Cherry-Pick][CI] Add check trigger and logic(#5191)

Modifications

Usage or Command

Accuracy Tests

Checklist

• Add at least a tag in the PR title.
  • Tag list: [[FDConfig],[APIServer],[Engine], [Scheduler], [PD Disaggregation], [Executor], [Graph Optimization], [Speculative Decoding], [RL], [Models], [Quantization], [Loader], [OP], [KVCache], [DataProcessor], [BugFix], [Docs], [CI], [Optimization], [Feature], [Benchmark], [Others], [XPU], [HPU], [GCU], [DCU], [Iluvatar], [Metax]]
  • You can add new tags based on the PR content, but the semantics must be clear.
• Format your code, run pre-commit before commit.
• Add unit tests. Please write the reason in this PR if no unit tests.
• Provide accuracy results.
• If the current PR is submitting to the release branch, make sure the PR has been submitted to the develop branch, then cherry-pick it to the release branch with the [Cherry-Pick] PR tag.

[paddle-bot]

Thanks for your contribution!

[codecov-commenter]

Codecov Report

❌ Patch coverage is 0% with 1 line in your changes missing coverage. Please review.
⚠️ Please upload report for BASE (develop@734fbcf). Learn more about missing BASE report.

Files with missing lines	Patch %	Lines
fastdeploy/model_executor/models/deepseek_v3.py	0.00%	1 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             develop    #7313   +/-   ##
==========================================
  Coverage           ?   74.14%           
==========================================
  Files              ?      383           
  Lines              ?    53631           
  Branches           ?     8411           
==========================================
  Hits               ?    39764           
  Misses             ?    11161           
  Partials           ?     2706           

Flag	Coverage Δ
GPU	74.14% <0.00%> (?)

Flags with carried forward coverage won't be shown. Click here to find out more.

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[fastdeploy-bot]

🤖 AI Code Review | 2026-04-11

📋 Review 摘要

PR 概述：DeepSeek V3 模型性能优化，包括 rotary kernel 支持 >65535 token、merge 算子支持多 head_dim

变更范围：custom_ops/gpu_ops/、model_executor/models/、tests/operators/

影响面 Tag：[Optimization] [OP] [Models]

📝 PR 规范检查

PR 标题和描述符合规范，包含 [Optimization] 和 [OP]、[Models] 标签，并填写了 Motivation 和 Modifications。

发现的问题

级别	文件	概述
🟡 建议	tests/operators/test_fused_rotary_position_encoding.py:118	test_large_num_tokens 只验证不抛出异常，没有验证输出正确性
🟡 建议	fastdeploy/model_executor/models/deepseek_v3.py:445	建议确认 max_token=1 硬编码值是否适用于所有场景（见评论详情）

总体评价

修改逻辑正确：rotary kernel 使用 2D grid 突破 65535 限制、merge 算子支持多 head_dim 的实现合理。但建议补充 merge_prefill_decode_output 的单元测试，并完善 test_large_num_tokens 的正确性验证。

[fastdeploy-bot]

🟡 建议 test_large_num_tokens 只验证算子不抛出异常，没有验证输出结果的正确性。

建议添加正确性验证（类似其他测试用例使用 _check_correctness）：

def test_large_num_tokens(self):
    num_tokens, num_heads, head_size = 65537, 1, 4
    num_kv_heads, rot_dim = 1, 4
    query_np = np.random.rand(num_tokens, num_heads, head_size).astype("float32")
    key_np = np.random.rand(num_tokens, num_kv_heads, head_size).astype("float32")
    position_ids_np = np.arange(num_tokens, dtype="int32")
    cos_sin_cache_np = self._make_cos_sin_cache(num_tokens, rot_dim)
    
    query_out, key_out = self._run_op(
        query_np, key_np, position_ids_np, cos_sin_cache_np, head_size, is_neox=False
    )
    # 添加正确性验证
    query_ref, key_ref = self._ref_rotary(
        query_np, key_np, position_ids_np, cos_sin_cache_np, head_size, is_neox=False
    )
    np.testing.assert_allclose(query_out, query_ref, rtol=1e-5, atol=1e-6)
    np.testing.assert_allclose(key_out, key_ref, rtol=1e-5, atol=1e-6)

[fastdeploy-bot]

❓ 疑问 max_token=1 被硬编码传递给 merge_prefill_decode_output。

在 CUDA kernel 中，max_token 用于计算 grid 的 z 维度：

const int warps = head_dim / 32;
const int tokens_block = (max_token + warps - 1) / warps;
dim3 grid_dims(batch_size, head_num, tokens_block);

当 max_token=1 时，tokens_block=1，token_id = warp_id（范围 0-31）。如果 seq_lens_this_time[bidb] > warps，会导致某些 token 无法被处理。

需要确认：

1. 在 DeepSeek V3 的使用场景中，seq_lens_this_time[bidb] 是否总是 ≤ warps？
2. 是否有场景需要一次解码生成多个 token（如 speculative decoding）？

建议：如果 max_token=1 是针对当前场景的优化，请添加注释说明原因；否则考虑使用 forward_meta.max_len_tensor_cpu[2]（max_dec_len_this_time）作为参数。