Performance Query: Is ~9.5 tokens/s normal for Kimi-K2-Thinking-Native on Xeon 6960P (MRDIMM 8800) + RTX 6000 Blackwell?

### Reminder

- [x] I have read the above rules and searched the existing issues.

### System Info

- ktransformers version: main branch
- sglang version: kvcache-ai/sglang, branch kimi_k2
- CPU: Single Intel Xeon 6960P (Running on a dual-socket machine, using 1 CPU)
- RAM: 12 * 96 GB DDR5 MRDIMM-8800
- GPU: Single RTX PRO 6000 Blackwell Server GPU

`nvidia-smi` output:
```
+-----------------------------------------------------------------------------------------+
| NVIDIA-SMI 580.119.02 Driver Version: 580.119.02 CUDA Version: 13.0 |
+-----------------------------------------+------------------------+----------------------+
| GPU Name Persistence-M | Bus-Id Disp.A | Volatile Uncorr. ECC |
| Fan Temp Perf Pwr:Usage/Cap | Memory-Usage | GPU-Util Compute M. |
| | | MIG M. |
|=========================================+========================+======================|
| 0 NVIDIA RTX PRO 6000 Blac... Off | 00000000:15:00.0 Off | 0 |
| N/A 43C P0 174W / 600W | 82771MiB / 97887MiB | 23% Default |
| | | Disabled |
+-----------------------------------------+------------------------+----------------------+

+-----------------------------------------------------------------------------------------+
| Processes: |
| GPU GI CI PID Type Process name GPU Memory |
| ID ID Usage |
|=========================================================================================|
| 0 N/A N/A 2332 G /usr/lib/xorg/Xorg 4MiB |
| 0 N/A N/A 5242 C sglang::scheduler 82730MiB |
+-----------------------------------------------------------------------------------------+
```
`lscpu` output:
```
Architecture: x86_64
CPU op-mode(s): 32-bit, 64-bit
Address sizes: 52 bits physical, 57 bits virtual
Byte Order: Little Endian
CPU(s): 144
On-line CPU(s) list: 0-143
Vendor ID: GenuineIntel
Model name: Intel(R) Xeon(R) 6960P
CPU family: 6
Model: 173
Thread(s) per core: 1
Core(s) per socket: 72
Socket(s): 2
Stepping: 1
CPU(s) scaling MHz: 36%
CPU max MHz: 3900.0000
CPU min MHz: 800.0000
BogoMIPS: 5400.00
Flags: fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush dts acpi mmx fxsr sse ss
e2 ss ht tm pbe syscall nx pdpe1gb rdtscp lm constant_tsc art arch_perfmon pebs bts rep_good nopl xtopol
ogy nonstop_tsc cpuid aperfmperf tsc_known_freq pni pclmulqdq dtes64 monitor ds_cpl vmx smx est tm2 ssse
3 sdbg fma cx16 xtpr pdcm pcid dca sse4_1 sse4_2 x2apic movbe popcnt tsc_deadline_timer aes xsave avx f1
6c rdrand lahf_lm abm 3dnowprefetch cpuid_fault epb cat_l3 cat_l2 cdp_l3 invpcid_single intel_ppin cdp_l
2 ssbd mba ibrs ibpb stibp ibrs_enhanced tpr_shadow vnmi flexpriority ept vpid ept_ad fsgsbase tsc_adjus
t bmi1 avx2 smep bmi2 erms invpcid cqm rdt_a avx512f avx512dq rdseed adx smap avx512ifma clflushopt clwb
intel_pt avx512cd sha_ni avx512bw avx512vl xsaveopt xsavec xgetbv1 xsaves cqm_llc cqm_occup_llc cqm_mbm
_total cqm_mbm_local avx_vnni avx512_bf16 wbnoinvd dtherm ida arat pln pts hwp hwp_act_window hwp_epp hw
p_pkg_req hfi avx512vbmi umip pku ospke waitpkg avx512_vbmi2 gfni vaes vpclmulqdq avx512_vnni avx512_bit
alg tme avx512_vpopcntdq la57 rdpid bus_lock_detect cldemote movdiri movdir64b enqcmd fsrm md_clear seri
alize tsxldtrk pconfig arch_lbr ibt amx_bf16 avx512_fp16 amx_tile amx_int8 flush_l1d arch_capabilities i
bpb_exit_to_user
Virtualization features:
Virtualization: VT-x
Caches (sum of all):
L1d: 6.8 MiB (144 instances)
L1i: 9 MiB (144 instances)
L2: 288 MiB (144 instances)
L3: 864 MiB (2 instances)
NUMA:
NUMA node(s): 2
NUMA node0 CPU(s): 0-71
NUMA node1 CPU(s): 72-143
Vulnerabilities:
Gather data sampling: Not affected
Indirect target selection: Not affected
Itlb multihit: Not affected
L1tf: Not affected
Mds: Not affected
Meltdown: Not affected
Mmio stale data: Not affected
Reg file data sampling: Not affected
Retbleed: Not affected
Spec rstack overflow: Not affected
Spec store bypass: Mitigation; Speculative Store Bypass disabled via prctl
Spectre v1: Mitigation; usercopy/swapgs barriers and __user pointer sanitization
Spectre v2: Mitigation; Enhanced / Automatic IBRS; IBPB conditional; PBRSB-eIBRS Not affected; BHI BHI_DIS_S
Srbds: Not affected
Tsa: Not affected
Tsx async abort: Not affected
Vmscape: Mitigation; IBPB before exit to userspace
```
`htop` output:
```

0[100] 9[100] 18[100] 27[100] 36[100] 45[100] 54[100] 63[100] 72[2.0] 81[8.4] 90[0.0] 99[0.0]108[0.0]117[0.0]126[100]135[0.0]
1[100] 10[100] 19[100] 28[100] 37[100] 46[100] 55[100] 64[100] 73[0.6] 82[0.0] 91[0.0]100[0.0]109[0.0]118[0.0]127[0.0]136[0.0]
2[100] 11[100] 20[100] 29[100] 38[100] 47[100] 56[100] 65[100] 74[0.0] 83[0.0] 92[0.0]101[0.0]110[0.0]119[0.0]128[0.0]137[0.0]
3[100] 12[100] 21[100] 30[100] 39[100] 48[100] 57[100] 66[100] 75[3.2] 84[0.0] 93[0.0]102[0.0]111[0.0]120[0.0]129[0.0]138[0.0]
4[100] 13[100] 22[100] 31[100] 40[100] 49[100] 58[100] 67[100] 76[1.9] 85[0.0] 94[0.0]103[0.0]112[0.0]121[0.0]130[0.0]139[0.0]
5[100] 14[100] 23[100] 32[100] 41[100] 50[100] 59[100] 68[100] 77[11.] 86[49.] 95[0.0]104[0.0]113[0.0]122[0.0]131[0.0]140[0.0]
6[100] 15[100] 24[100] 33[100] 42[100] 51[100] 60[100] 69[100] 78[0.6] 87[0.0] 96[0.0]105[0.0]114[0.0]123[0.0]132[0.0]141[0.0]
7[100] 16[100] 25[100] 34[100] 43[100] 52[100] 61[100] 70[100] 79[0.0] 88[0.0] 97[0.0]106[0.0]115[0.0]124[0.0]133[2.0]142[0.0]
8[100] 17[100] 26[100] 35[100] 44[100] 53[100] 62[100] 71[100] 80[10.] 89[0.0] 98[0.0]107[0.0]116[0.0]125[0.0]134[0.0]143[0.6]
Mem[|||||||||||||||||||||||||||||||||||||||||||||||||604G/1.10T] Load average: 68.00 39.29 22.10
Swp[||| 20.7M/977M] Uptime: 00:23:53

[Main] [I/O]
PID USER PRI NI VIRT RES SHR S CPU%▽MEM% TIME+ Command
5242 admin 20 0 1519G 1106G 512G S 7365.5 97.9 5h41:56 sglang::scheduler
5881 admin 20 0 1519G 1106G 512G R 100.2 97.9 4:17.12 sglang::scheduler
5882 admin 20 0 1519G 1106G 512G R 100.2 97.9 4:21.58 sglang::scheduler
5883 admin 20 0 1519G 1106G 512G R 100.2 97.9 4:21.08 sglang::scheduler
5889 admin 20 0 1519G 1106G 512G R 100.2 97.9 4:20.86 sglang::scheduler
5890 admin 20 0 1519G 1106G 512G R 100.2 97.9 4:21.45 sglang::scheduler
5891 admin 20 0 1519G 1106G 512G R 100.2 97.9 4:19.52 sglang::scheduler
5893 admin 20 0 1519G 1106G 512G R 100.2 97.9 4:19.65 sglang::scheduler
5894 admin 20 0 1519G 1106G 512G R 100.2 97.9 4:21.27 sglang::scheduler
5896 admin 20 0 1519G 1106G 512G R 100.2 97.9 4:22.14 sglang::scheduler
5897 admin 20 0 1519G 1106G 512G R 100.2 97.9 4:20.76 sglang::scheduler
5899 admin 20 0 1519G 1106G 512G R 100.2 97.9 4:20.11 sglang::scheduler
5900 admin 20 0 1519G 1106G 512G R 100.2 97.9 4:22.11 sglang::scheduler
5905 admin 20 0 1519G 1106G 512G R 100.2 97.9 4:19.88 sglang::scheduler
5906 admin 20 0 1519G 1106G 512G R 100.2 97.9 4:20.11 sglang::scheduler
5907 admin 20 0 1519G 1106G 512G R 100.2 97.9 4:17.18 sglang::scheduler
5908 admin 20 0 1519G 1106G 512G R 100.2 97.9 4:19.98 sglang::scheduler
5910 admin 20 0 1519G 1106G 512G R 100.2 97.9 4:19.49 sglang::scheduler
5912 admin 20 0 1519G 1106G 512G R 100.2 97.9 4:20.80 sglang::scheduler
5913 admin 20 0 1519G 1106G 512G R 100.2 97.9 4:20.94 sglang::scheduler
5914 admin 20 0 1519G 1106G 512G R 100.2 97.9 4:19.36 sglang::scheduler
5915 admin 20 0 1519G 1106G 512G R 100.2 97.9 4:19.60 sglang::scheduler
5917 admin 20 0 1519G 1106G 512G R 100.2 97.9 4:20.91 sglang::scheduler
5921 admin 20 0 1519G 1106G 512G R 100.2 97.9 4:19.92 sglang::scheduler
5922 admin 20 0 1519G 1106G 512G R 100.2 97.9 4:18.53 sglang::scheduler
5924 admin 20 0 1519G 1106G 512G R 100.2 97.9 4:18.44 sglang::scheduler
5925 admin 20 0 1519G 1106G 512G R 100.2 97.9 4:19.54 sglang::scheduler
5926 admin 20 0 1519G 1106G 512G R 100.2 97.9 4:19.56 sglang::scheduler
5927 admin 20 0 1519G 1106G 512G R 100.2 97.9 4:17.98 sglang::scheduler
5931 admin 20 0 1519G 1106G 512G R 100.2 97.9 4:20.18 sglang::scheduler
5932 admin 20 0 1519G 1106G 512G R 100.2 97.9 4:18.45 sglang::scheduler
5937 admin 20 0 1519G 1106G 512G R 100.2 97.9 4:20.55 sglang::scheduler
5938 admin 20 0 1519G 1106G 512G R 100.2 97.9 4:19.47 sglang::scheduler
5940 admin 20 0 1519G 1106G 512G R 100.2 97.9 4:19.62 sglang::scheduler
5942 admin 20 0 1519G 1106G 512G R 100.2 97.9 4:20.95 sglang::scheduler
5943 admin 20 0 1519G 1106G 512G R 100.2 97.9 4:19.16 sglang::scheduler
5944 admin 20 0 1519G 1106G 512G R 100.2 97.9 4:20.44 sglang::scheduler
5945 admin 20 0 1519G 1106G 512G R 100.2 97.9 4:19.75 sglang::scheduler
5946 admin 20 0 1519G 1106G 512G R 100.2 97.9 4:16.58 sglang::scheduler
5947 admin 20 0 1519G 1106G 512G R 100.2 97.9 4:20.47 sglang::scheduler
5948 admin 20 0 1519G 1106G 512G R 100.2 97.9 4:20.74 sglang::scheduler
5949 admin 20 0 1519G 1106G 512G R 100.2 97.9 4:20.32 sglang::scheduler
5950 admin 20 0 1519G 1106G 512G R 100.2 97.9 4:19.81 sglang::scheduler
5957 admin 20 0 1519G 1106G 512G R 100.2 97.9 3:14.38 sglang::scheduler
5884 admin 20 0 1519G 1106G 512G R 99.6 97.9 4:19.12 sglang::scheduler
5885 admin 20 0 1519G 1106G 512G R 99.6 97.9 4:22.27 sglang::scheduler
5886 admin 20 0 1519G 1106G 512G R 99.6 97.9 4:23.13 sglang::scheduler
5887 admin 20 0 1519G 1106G 512G R 99.6 97.9 4:21.88 sglang::scheduler
5888 admin 20 0 1519G 1106G 512G R 99.6 97.9 4:22.35 sglang::scheduler
5892 admin 20 0 1519G 1106G 512G R 99.6 97.9 4:18.73 sglang::scheduler
5895 admin 20 0 1519G 1106G 512G R 99.6 97.9 4:19.71 sglang::scheduler
5898 admin 20 0 1519G 1106G 512G R 99.6 97.9 4:21.37 sglang::scheduler
5901 admin 20 0 1519G 1106G 512G R 99.6 97.9 4:22.38 sglang::scheduler

```
`pqos -m` output:
```
CORE IPC MISSES MBL[MB/s] MBR[MB/s]
0 1.55 31047k 2313.1 0.0
1 1.19 32127k 2263.4 0.0
2 1.19 32470k 2280.2 0.8
3 1.20 32153k 2192.9 0.0
4 1.19 32516k 2290.1 0.0
5 1.20 32278k 2202.6 0.0
6 1.19 32482k 2250.1 0.6
7 1.18 33695k 2345.2 0.0
8 1.19 33304k 2322.0 0.2
9 1.19 33102k 2333.8 0.0
10 1.19 33330k 2279.6 1.5
11 1.18 33854k 2301.1 0.0
12 1.18 33711k 2378.3 0.0
13 1.18 33701k 2366.7 0.2
14 1.17 33821k 2352.6 0.0
15 1.18 33426k 2323.1 0.0
16 1.32 25082k 1736.0 0.0
17 1.33 24887k 1742.8 0.0
18 1.32 24973k 1729.5 0.0
19 1.33 24732k 1715.3 0.2
20 1.33 24418k 1723.8 0.0
21 1.33 24841k 1761.1 0.0
22 1.34 24200k 1728.2 0.6
23 1.33 24969k 1719.6 1.5
24 1.33 25046k 1734.1 0.0
25 1.35 24286k 1657.3 0.0
26 1.36 23781k 1659.0 0.6
27 1.36 23526k 1611.6 0.0
28 1.40 20653k 1441.3 0.0
29 1.40 20607k 1404.6 0.0
30 1.42 19958k 1398.5 0.0
```

### Reproduction

I am following the tutorial: `Running Kimi-K2-Thinking with SGLang and KT-Kernel`.
Launch Command:
```
python -m sglang.launch_server \
--host 0.0.0.0 \
--port 8000 \
--model /mnt/m2b/models/moonshotai/Kimi-K2-Thinking \
--kt-weight-path /mnt/m2b/models/moonshotai/Kimi-K2-Thinking \
--kt-cpuinfer 72 \
--kt-threadpool-count 1 \
--kt-num-gpu-experts 32 \
--kt-method RAWINT4 \
--kt-gpu-prefill-token-threshold 400 \
--kt-max-deferred-experts-per-token 1 \
--trust-remote-code \
--mem-fraction-static 0.94 \
--served-model-name Kimi-K2-Thinking \
--enable-mixed-chunk \
--tensor-parallel-size 1 \
--enable-p2p-check \
--disable-shared-experts-fusion \
--chunked-prefill-size 16384 \
--max-total-tokens 16384 \
--max-running-requests 1 \
--attention-backend flashinfer
```
Output Logs (showing generation speed) attached:
```
[2025-12-17 23:36:34] Decode batch, #running-req: 1, #token: 2778, token usage: 0.17, cuda graph: True, gen throughput (token/s): 0.14, #queue-req: 0,
[2025-12-17 23:36:38] Decode batch, #running-req: 1, #token: 2818, token usage: 0.17, cuda graph: True, gen throughput (token/s): 9.53, #queue-req: 0,
[2025-12-17 23:36:42] Decode batch, #running-req: 1, #token: 2858, token usage: 0.17, cuda graph: True, gen throughput (token/s): 9.54, #queue-req: 0,
[2025-12-17 23:36:47] Decode batch, #running-req: 1, #token: 2898, token usage: 0.18, cuda graph: True, gen throughput (token/s): 9.53, #queue-req: 0,
[2025-12-17 23:36:51] Decode batch, #running-req: 1, #token: 2938, token usage: 0.18, cuda graph: True, gen throughput (token/s): 9.53, #queue-req: 0,
[2025-12-17 23:36:55] Decode batch, #running-req: 1, #token: 2978, token usage: 0.18, cuda graph: True, gen throughput (token/s): 9.53, #queue-req: 0,
[2025-12-17 23:36:59] Decode batch, #running-req: 1, #token: 3018, token usage: 0.18, cuda graph: True, gen throughput (token/s): 9.53, #queue-req: 0,
```

[sglang.txt](https://github.com/user-attachments/files/24227217/sglang.txt)

### Others

I am observing a decoding speed of approximately 9.5 tokens/s.
Given the hardware specifications:

- CPU: Intel Xeon 6960P with AMX support.
- Memory: High-bandwidth MRDIMM-8800 (12 channels populated).
- GPU: RTX PRO 6000 Blackwell.

I would like to ask if this speed is within the expected range for this configuration using the kimi_k2 branch and KT-Kernel?
Considering the high memory bandwidth of the MRDIMMs and the capabilities of the CPU/GPU, I wanted to confirm if there are potential bottlenecks I should look into (e.g., NUMA configuration on the dual-socket board despite using one CPU, specific kernel flags, or quantization overhead).