⚙ SGLang Diffusion Is Retiring torch.compile, With Help From an AI Agent After testing 35 diffusion model presets on H100 and H200 GPUs, SGLang Diffusion found that eager execution now usually matches or outperforms torch.compile. Zhihu contributor BBuf argues that this is more than a benchmark result. Agent-written optimizations and Breakable CUDA Graphs can now replace the two main benefits that torch.compile once provided. 1️⃣ The “seven sins” are three production failures torch.compile can fuse kernels and reduce Python overhead. But in production diffusion serving, BBuf found that its operational cost often outweighed those gains. 🔹 Slow and unpredictable Compilation may take several minutes, and complex models can take more than half an hour. Performance also changes unpredictably across compiler modes, models, and GPU generations. An optimization that works on one GPU may become a regression on another. 🔹 Hostile to manual optimization A custom fused kernel may accelerate eager execution but slow down the compiled model. Graph breaks or broader compiler fusion can erase the improvement. Developers then have to debug both their kernels and the compiler’s changing graph decisions. 🔹 Opaque and difficult to maintain Compiled traces are hard to map back to Python calls. Generated kernels are also difficult to inspect, benchmark independently, or modify. Over time, engineers are pulled into compiler-specific fusion rules and workarounds instead of optimizing the actual model. 2️⃣ The current benchmark changed the calculation In SGLang Diffusion’s latest H100 and H200 tests, torch.compile was usually slower than eager execution or roughly tied with it. Compile jobs taking longer than three minutes were stopped. Several models also encountered timeouts or runtime failures. Some models had previously benefited from compilation. But after explicit kernel optimization, most of those gaps disappeared. For models dominated by launch overhead, eager execution plus Breakable CUDA Graphs could match or beat the compiled path. 3️⃣ An Agent replaced the hidden optimizations The team did not build an elaborate multi-agent system. They used Fable with a simple optimization flow and two Kernel Design Agent skills. The Agent profiled the VAE and model components, identified bottlenecks, implemented explicit optimizations, and repeatedly checked correctness and performance. Within two weeks, this workflow optimized more than 30 diffusion models. Their eager paths reached or surpassed the performance previously provided by torch.compile. The crucial difference is ownership: these optimizations now exist as visible, testable, and maintainable code. 4️⃣ Breakable CUDA Graph replaced the other benefit Kernel fusion is only one reason to use torch.compile. The other is reducing framework overhead through graph execution. SGLang previously relied on compiler-assisted Piecewise CUDA Graphs. After dealing with substantial compiler complexity, it developed Breakable CUDA Graph, which captures reusable execution paths while allowing controlled graph breaks. SGLang Diffusion now uses the same approach. It is easier to integrate into real serving systems and can outperform the compiled path on overhead-heavy models. 5️⃣ The deeper shift is from compiler magic to Agent-owned code A general compiler applies hidden transformations at runtime. An Agent can instead inspect the workload, reproduce useful transformations as source code, and validate them on the target hardware. This makes performance work easier to profile, review, debug, and adapt. BBuf’s claim is specifically about diffusion serving, not every PyTorch workload. But within that scope, the conclusion is blunt: If Agents can reproduce the useful kernel optimizations, while CUDA Graph techniques remove framework overhead, torch.compile no longer has a unique role. 🔗 Full Reading: 🔗 Performance results: 🔗 Kernel Design Agents: 🔗 SGLang: 🔗 Agent optimization skills: #SGLang# #PyTorch# #CUDA# #AIInfra# #DiffusionModels# #AIAgents#

Tracking and managing inventory has never been easier. Gemini Enterprise can connect different systems to give you a unified view of your stock. It can also help you spot stock risks before they impact your sales and customers. Try it yourself →

My colleague wanted to sell his car to offset his ₦1.8m house rent. Downgrade to something smaller. His wife said she would buy the car. “You will buy my car? Where did you see money?” he asked her. She said, “My brother raised ₦6 million for me last year to start a business. I fixed the money.” “And I wasn’t aware? You still collect ₦100k from me every month to take care of yourself?” It’s been two weeks now since he left the house. He said he is no longer interested in the marriage.

Those who have been creating original content, building their pages properly, and engaging with fellow small accounts before now will have little or no issues meeting the requirements of the new monetisation programme. Those who have been camping under big accounts "only" for the sake of getting 5M impressions before now, while never engaging with fellow small accounts or trying to create their own content, will now have to work three times as hard to get monetised.

🧩 Kimi K3’s MoE and Attention Are Built Around Trade-offs, Not Tricks Kimi K3’s open release has drawn attention to its scale. But its architecture tells a more useful story: the hardest part of scaling is keeping quality, efficiency, and stability in balance. Zhihu contributor 苏剑林 @Jianlin_S explains the design logic behind two core components: Stable LatentMoE and K3’s hybrid attention. At a high level: K3 = KDA + MLA + Stable LatentMoE + AttnRes 1️⃣ Stable LatentMoE: more experts at similar cost LatentMoE compresses each token into a smaller latent space before routing it to experts, then projects the result back to the full hidden dimension. This reduces expert computation and communication. The saved budget can support more, narrower experts without greatly increasing training or inference cost. But the longer projection chain also magnifies numerical instability. K3 introduces three fixes. 🔹 SiTU-GLU softly caps extreme activations in both branches of the expert network. Compared with hard clipping, soft capping preserves smoother optimization. 🔹 RMSNorm is placed before the final up-projection. It stabilizes training and helps balance routed experts against shared experts. 🔹 Quantile Balancing replaces the previous load-balancing update, which became unreliable as the expert pool grew. It approximates global routing quantiles with histograms, allowing efficient aggregation across machines. The broader lesson is clear: scaling MoE is not just about adding experts. Routing, activation ranges, normalization, and distributed communication must scale with them. 2️⃣ Why K3 still uses MLA Some newer models have moved away from MLA, partly because speculative decoding changes the inference trade-off. MLA keeps KV Cache small and remains highly competitive under fixed training and memory budgets. But its decoding path is relatively compute-heavy, leaving less room for Multi-Token Prediction to trade extra computation for speed. Other attention designs simply move the bottleneck: 🔹 Smaller designs may reduce computation but lose quality or require a larger KV Cache. 🔹 Larger designs can recover quality, but increase training and prefill costs. An ideal replacement would preserve quality, reduce KV Cache, lower decoding compute, and cost no more during training or prefill. No simple design currently satisfies all four conditions. K3 therefore keeps MLA and combines it with KDA. The linear-attention layers handle most long-context processing efficiently, while MLA preserves full-attention capacity where it matters. 3️⃣ “Abandoning MLA” is not so simple Architectures that appear to replace MLA may still retain its core intuition. For example, a wide MQA design with shared K and V resembles MLA’s decoding form. Sparsity and compression can then reduce its compute and cache costs. This can work, but it introduces more infrastructure complexity. So the current debate is less about whether MLA is obsolete. It is about which combination of full, linear, sparse, and compressed attention offers the best system-level trade-off. 4️⃣ Why K3 can remove RoPE K3 removes RoPE from its MLA layers. That would hurt a pure-MLA model. But K3 is a hybrid of KDA and MLA. KDA’s DeltaNet-style updates already introduce an implicit positional transformation. In this sense, KDA provides something similar to a generalized form of RoPE for the full network. Adding explicit RoPE back produced little difference, so K3 followed the simpler design. K3 is not truly position-free. Its positional structure is partly carried by KDA instead of an explicit embedding. ⚙ The real architecture lesson None of these choices is especially flashy in isolation. Stable LatentMoE controls the numerical and routing problems created by more experts. KDA and MLA divide long-context work according to their strengths. NoPE removes a redundant component only after the hybrid architecture makes it unnecessary. K3’s main design principle is therefore not novelty for its own sake. Every architectural change must justify itself across quality, efficiency, and stability. 🔗 Full reading: 📖Blog post: #KimiK3# #MoE# #Attention# #LLM# #AIInfra# #OpenSourceAI#

Build _______ with Gemini. Fill in the blank at #BuildwithGemini#, a complimentary, hands-on workshop on building, scaling, and deploying secure, production-ready AI agents. Find an event near you →