yushengsu/rope-dump-DeepSeek-V3.1-Base

RoPE (Rotary Position Embedding) Input/Output Dump

Dump of the apply_rope_inplace kernel from SGLang, captured during inference with DeepSeek-V3.1-Base.

Setup

• Model: deepseek-ai/DeepSeek-V3.1-Base (FP8, MLA architecture)
• Framework: SGLang (sglang-private, commit 01e9b55 — "Add LoRA MoE support with fused Triton kernels and TP slicing")
• Attention Backend: trtllm_mla (default on sm100 / GB300)
• TP: 4 (4x NVIDIA GB300)
• CUDA Graph: disabled
• Input Prompt: "hello, how are you?" (7 tokens after tokenization)

Server Startup

bash start_server_deepseek.sh

See start_server_deepseek.sh for the full launch command.

Function Signature

# sglang-private/python/sglang/jit_kernel/rope.py

@register_custom_op(mutates_args=["q", "k"])
def apply_rope_inplace(
    q: torch.Tensor,         # [num_tokens, num_qo_heads, rope_dim] — modified in-place
    k: torch.Tensor,         # [num_tokens, num_kv_heads, rope_dim] — modified in-place
    cos_sin_cache: torch.Tensor,  # [max_position, rope_dim], float32
    positions: torch.Tensor,      # [num_tokens], int64
    *,
    is_neox: bool,           # False for DeepSeek (GPT-J interleaved style)
    rope_dim: int = 0,       # 64
) -> None

Dump Files

Tensors are stored in safetensors format (rope_dump.safetensors).

Tensor Keys in rope_dump.safetensors

Key	Shape	Dtype	Description
q_input	[7, 32, 64]	bfloat16	Query tensor before RoPE (7 tokens, 32 qo_heads on TP0, rope_dim=64)
q_output	[7, 32, 64]	bfloat16	Query tensor after RoPE
k_input	[7, 1, 64]	bfloat16	Key tensor before RoPE (7 tokens, 1 kv_head, rope_dim=64)
k_output	[7, 1, 64]	bfloat16	Key tensor after RoPE
cos_sin_cache	[164096, 64]	float32	Precomputed cos/sin cache. First 32 cols = cos, last 32 cols = sin
positions	[7]	int64	Position indices: [0, 1, 2, 3, 4, 5, 6]

Other Files

File	Description
call_0_meta.json	Shapes, dtypes, is_neox, rope_dim
start_server_deepseek.sh	SGLang server startup script

Parameters

is_neox = False  (GPT-J interleaved style)
rope_dim = 64
positions = [0, 1, 2, 3, 4, 5, 6]

Quick Load Example

from safetensors.torch import load_file

tensors = load_file("rope_dump.safetensors")

q_in  = tensors["q_input"]       # [7, 32, 64] bfloat16
q_out = tensors["q_output"]      # [7, 32, 64] bfloat16
k_in  = tensors["k_input"]       # [7, 1, 64]  bfloat16
k_out = tensors["k_output"]      # [7, 1, 64]  bfloat16
cos_sin_cache = tensors["cos_sin_cache"]  # [164096, 64] float32
positions     = tensors["positions"]      # [7] int64

print(f"Q: {q_in.shape} -> {q_out.shape}")  # [7, 32, 64]
print(f"K: {k_in.shape} -> {k_out.shape}")  # [7, 1, 64]
print(f"positions: {positions.tolist()}")     # [0, 1, 2, 3, 4, 5, 6]

Load from HuggingFace Hub

from huggingface_hub import hf_hub_download
from safetensors.torch import load_file

path = hf_hub_download(
    repo_id="yushengsu/rope-dump-DeepSeek-V3.1-Base",
    filename="rope_dump.safetensors",
    repo_type="dataset",
)
tensors = load_file(path)
print(list(tensors.keys()))
# ['cos_sin_cache', 'k_input', 'k_output', 'positions', 'q_input', 'q_output']

Notes

• At position=0, cos=1 and sin=0, so k_output == k_input (identity rotation).
• Q changes even at position 0 because MLA's q_b_proj mixes head information before RoPE.
• This dump captures the first layer's RoPE call during prefill of the prompt.
• During decode steps, DeepSeek's trtllm_mla backend fuses RoPE into the attention kernel, so apply_rope_inplace is only called during prefill.