TeleFuser: Skill for Claude Code

Optimize a Pipeline

Require a correct, reproducible baseline before changing performance behavior. Read current implementations and canonical docs rather than relying on hard-coded speedup or memory estimates.

Define the target

Record the workload, model/checkpoint, task, shape, frame count, step count, dtype, attention backend, devices, warmup, measured latency/throughput, peak memory, and quality or parity criterion. Distinguish latency, throughput, capacity, and output cadence; they require different choices.

If no baseline exists, profile first. Use .claude/skills/profile-pipeline/SKILL.md and docs/en/profiler.md.

Reuse supported mechanisms

Inspect the closest model and pipeline before choosing an optimization. Consult the relevant current docs:

• docs/en/ops.md and docs/en/attention.md
• docs/en/parallel.md
• docs/en/offload.md
• docs/en/feature_cache.md
• docs/en/torch_compile_compatibility.md
• docs/en/configuration.md

Apply these constraints:

• Route model operations through telefuser.ops; do not import Triton kernels directly from models/.
• Preserve exact semantics when replacing an op: layout, normalization, RoPE, masks, causal behavior, scale, dtype, and numerical tolerances must match.
• Use current AttentionConfig, ModelRuntimeConfig, ParallelConfig, FeatureCacheConfig, CompileConfig, QuantConfig, and OffloadConfig APIs as implemented in the repository.
• Confirm that the target model and stage implement the selected parallel or optimization path. A config field existing does not prove model support.
• Treat sparse attention, feature caching, quantization, distillation, and approximate computation as behavior or quality changes; require explicit user agreement and parity/quality evidence.
• Do not add a new public interface, configuration system, environment variable, loader, or parallel abstraction as part of optimization without first demonstrating a missing extension point and obtaining approval.