菜单
Accelerate LLM inference 6x by using block diffusion for parallel token drafting with tight coupling to the target model's hidden representations, achieving higher speedups than existing speculative methods without quality loss.
默认使用会先检查来源的 Prompt；你也可以切换为直接命令，或下载本地副本。
决定是否安装前，请先阅读 SKILL.md，以及 SkillsMP 当前展示的配套文件。
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
直接命令不会经过审查 Prompt；运行前请先检查来源。
npx skills add https://github.com/ADu2021/skillXiv --skill dflash-block-diffusion-speculative-decoding命令会保持在同一行。复制前请横向滚动并检查完整内容。
想先保存到本地？可下载 SkillsMP 当前能够提供的文件。
Convert arXiv papers into ready-to-use agent skills using category-aware extraction. First classifies the paper into one or more of 11 research categories, then applies a specialized extraction pipeline for each category — because different types of papers produce different types of usable knowledge. A single paper can yield multiple skills if it spans categories. Use this skill whenever the user wants to turn a paper into a skill, extract practical techniques from research, build a skill library from papers, convert arXiv papers into reusable agent instructions, or batch-process multiple papers into skills. Also trigger when someone asks about extracting actionable knowledge from papers, making research practical for LLM agents, or systematically converting academic contributions into structured agent capabilities.
Establish regret bounds for behavior cloning with discretized actions combining statistical error and quantization error terms. Prove smoothness requirements for safe quantizer design, show that learning-based quantizers fail these requirements, and propose model-based augmentation to reduce error dependence from H² to H.
Dynamically allocate LoRA ranks per-layer during fine-tuning instead of using fixed uniform ranks. Learn optimal rank for each layer and subject via variational framework with discretized exponential distribution, reducing memory footprint while maintaining fidelity and text-alignment.
基于 SOC 职业分类
正在显示 SKILL.md
| name | dflash-block-diffusion-speculative-decoding |
| title | DFlash: Block Diffusion for Flash Speculative Decoding |
| version | 0.0.2 |
| engine | skillxiv-v0.0.2-claude-opus-4.6 |
| license | MIT |
| url | https://arxiv.org/abs/2602.06036 |
| keywords | ["Speculative Decoding","Diffusion Models","Parallel Generation","Inference Acceleration","Target Model Conditioning"] |
| description | Accelerate LLM inference 6x by using block diffusion for parallel token drafting with tight coupling to the target model's hidden representations, achieving higher speedups than existing speculative methods without quality loss. |
Autoregressive LLMs generate tokens sequentially, creating inference bottlenecks. Speculative decoding helps but is limited by reliance on autoregressive drafting—inherently sequential processes that cap speedups around 2-3x. Diffusion models enable parallel generation but typically underperform in quality. DFlash bridges this gap: parallel drafting with quality matching the target model.
DFlash combines [parallel block diffusion, target model conditioning, KV injection] to enable fast high-quality token drafting. The insight is "the target knows best"—leveraging frozen target model representations substantially improves draft quality without needing massive draft models.
Freeze the target LLM and extract intermediate layer hidden states to guide draft generation. Cache these for efficiency.
# Extract conditioning from target model
def extract_target_conditioning(target_model, input_ids, layer_indices):
"""
Forward target model; extract hidden states at specified layers.
Returns conditioning features for each layer.
"""
conditioning = {}
with torch.no_grad():
output = target_model(
input_ids,
output_hidden_states=True,
return_dict=True
)
for layer_idx in layer_indices:
# Extract and compress hidden state
h = output.hidden_states[layer_idx]
# Project to draft model's key-value dimension
conditioning[layer_idx] = project_to_kv(h)
conditioning
Inject target model context directly into draft model's Key-Value cache at every decoder layer, replacing standard feature fusion.
# KV injection into draft model cache
def inject_target_conditioning_to_kv(
draft_kv_cache, target_conditioning, layer_idx,
injection_strength=0.5
):
"""
Inject target model hidden states directly into KV cache.
Enables acceptance scaling with depth.
"""
k, v = draft_kv_cache[layer_idx]
# Target features as additional context in value
target_feat = target_conditioning[layer_idx]
# Blend target features with draft KV
v_augmented = v + injection_strength * target_feat
draft_kv_cache[layer_idx] = (k, v_augmented)
return draft_kv_cache
Initialize diffusion process with random anchor tokens at block boundaries. Use scheduled denoising to generate multiple tokens in parallel.
# Block diffusion generation
def block_diffusion_generate(
draft_model, context, block_size=4, num_steps=15
):
"""
Generate block_size tokens in parallel using diffusion.
Anchor tokens at block boundaries guide generation.
"""
batch_size = context.shape[0]
seq_len = context.shape[1]
# Initialize block: random anchors at boundaries
block_tokens = torch.randint(
0, draft_model.vocab_size,
(batch_size, block_size)
)
# Reverse diffusion: denoise tokens over num_steps
for step in range(num_steps):
noise_schedule = 1.0 - (step / num_steps)
# Predict token logits
logits = draft_model.denoise(
context, block_tokens,
timestep=step, num_steps=num_steps
)
# Sample new tokens with scheduled noise
block_tokens = sample_with_noise(
logits, block_tokens, noise_schedule
)
return block_tokens
During training, weight loss to prioritize early tokens in block, which are most predictable.
# Loss weighting for block diffusion training
def weighted_diffusion_loss(
predictions, targets, position_weights=None
):
"""
Weight tokens early in block more heavily.
Early tokens are easier to predict; later tokens need guidance.
"""
if position_weights is None:
# Default: exponential decay favoring early positions
block_size = targets.shape[-1]
position_weights = torch.exp(
-torch.arange(block_size, dtype=torch.float32) * 0.3
)
position_weights = position_weights / position_weights.sum()
# Per-position loss weighting
loss = F.cross_entropy(
predictions.reshape(-1, predictions.shape[-1]),
targets.reshape(-1),
reduction='none'
)
loss = loss.reshape(targets.shape)
# Apply position weighting
weighted_loss = (loss * position_weights).mean()
return weighted_loss
Generate drafts in parallel blocks; verify and accept valid tokens; fall back to target model only when needed.
# Speculative decoding with DFlash
def flash_speculative_decode(
target_model, draft_model, input_ids,
max_new_tokens=100, block_size=4
):
"""
Generate tokens using DFlash speculative decoding.
"""
generated = input_ids.clone()
target_conditioning = extract_target_conditioning(
target_model, input_ids, layer_indices=[12, 24]
)
while generated.shape[1] < input_ids.shape[1] + max_new_tokens:
# Draft: generate block in parallel
draft_block = block_diffusion_generate(
draft_model, generated, block_size=block_size
)
# Verify against target model
with torch.no_grad():
target_logits = target_model(
torch.cat([generated, draft_block], dim=1)
).logits
# Acceptance checking: compare probabilities
accepted = verify_tokens(
draft_block,
target_logits[:, -block_size:],
temperature=1.0
)
# Add verified tokens; fall back to target if none accepted
num_accepted = accepted.sum().item()
if num_accepted > 0:
generated = torch.cat([
generated,
draft_block[:, :num_accepted]
], dim=1)
else:
# Fall back to target model
next_token = target_model(generated).logits[:, -1].argmax(dim=-1)
generated = torch.cat([generated, next_token.unsqueeze(1)], dim=1)
return generated
When to use: High-throughput inference settings (serving, batch processing) where latency is critical. Less beneficial for interactive single-token-at-a-time scenarios.
Hyperparameters:
Common pitfalls:
Scaling: Speedup scales with block size and acceptance rate. Typical 4-6x speedup on standard benchmarks. Works best with strong target models and sufficient compute for parallel drafting.
Paper: https://arxiv.org/abs/2602.06036 Code: Available at author's repository Related work: Speculative decoding, EAGLE, diffusion-based generation, fast inference