[mmangkad]

Motivation

Three extra-b-test-4-gpu-b200 jobs fail on main (run 33962519343): test_kimi_linear_dcp4.py, test_kimi_linear_dcp_dspark4.py, test_unified_radix_cache_kl_dcp.py, all Kimi Linear on tokenspeed_mla + fp8_e4m3 under DCP. Every TP rank crashes during prefill and the server is SIGKILLed:

File ".../layers/attention/tokenspeed_mla_backend.py", line 265, in prepare_prefill_qkv
    cos_sin_cache=layer.rotary_emb.cos_sin_cache,
AttributeError: 'NoneType' object has no attribute 'cos_sin_cache'

Surfaced by #35770, which taught resolve_attn_backend() to unwrap HybridLinearAttnBackend.full_attn_backend. Kimi Linear is hybrid KDA/MLA, so prefill used to resolve to the wrapper, which has no prepare_prefill_qkv hook, and took the BF16 path guarded by if self.rotary_emb is not None. It now resolves to TokenspeedMLABackend, whose hook dereferences layer.rotary_emb unconditionally, and Kimi Linear builds its MLA layers with skip_rope=True, which leaves rotary_emb as None.

Modifications

_fused_rope_fp8_quantize gains a NoPE branch: when cos_sin_cache is None, keep the FP8 quantize and the packed [nope | pe] layout, drop only the rotation. It is an early return, so RoPE layers are unaffected.

The cos_sin_cache and is_neox_style reads become None checks rather than getattr defaults, and _fused_rope_fp8_quantize becomes a @staticmethod with a unit case, since nothing PR-gated covered this path.

Test Plan

Test	Result
dcp/test_kimi_linear_dcp4.py	OK, 3 tests
dcp/test_kimi_linear_dcp_dspark4.py	OK, GSM8K 0.905 (threshold 0.88)
radix_cache/unified_radix_tree/test_unified_radix_cache_kl_dcp.py	OK, avg_kl_div 0.0078 (threshold 0.01)
attention/unittests/mla/test_tokenspeed_mla.py	OK

test_kimi_linear_dcp4.py was also reproduced red by stashing the patch.

---

CI States

Latest PR Test (Base): ❌ Run #33984976309
Latest PR Test (Extra): ❌ Run #33984976221
Latest PR Test (AMD ROCm 7.2): ❌ Run #33984976307

[mmangkad]

/rerun-test test/registered/dcp/test_kimi_linear_dcp4.py test/registered/dcp/test_kimi_linear_dcp_dspark4.py test/registered/radix_cache/unified_radix_tree/test_unified_radix_cache_kl_dcp.py test/registered/attention/unittests/mla/test_tokenspeed_mla.py

[github-actions]

Results for /rerun-test test/registered/dcp/test_kimi_linear_dcp4.py test/registered/dcp/test_kimi_linear_dcp_dspark4.py test/registered/radix_cache/unified_radix_tree/test_unified_radix_cache_kl_dcp.py test/registered/attention/unittests/mla/test_tokenspeed_mla.py:

🚀 4-gpu-b200 (4 tests): ✅ View workflow run

cd test/ && python3 registered/dcp/test_kimi_linear_dcp4.py
cd test/ && python3 registered/dcp/test_kimi_linear_dcp_dspark4.py
cd test/ && python3 registered/radix_cache/unified_radix_tree/test_unified_radix_cache_kl_dcp.py
cd test/ && python3 registered/attention/unittests/mla/test_tokenspeed_mla.py

🚀 1-gpu-h100 (1 test): ✅ View workflow run

cd test/ && python3 registered/attention/unittests/mla/test_tokenspeed_mla.py