Zixi "Oz" Li PRO
OzTianlu
AI & ML interests
My research focuses on deep reasoning with small language models, Transformer architecture innovation, and knowledge distillation for efficient alignment and transfer.
Recent Activity
reacted to theirpost with 🤗 3 days ago
https://github.com/lizixi-0x2F/March
I just released March, an open-source high-performance KV cache sharing library for LLM inference that uses Trie-based prefix deduplication.
When you run LLM services, you often see thousands of requests sharing the same system prompt and conversation history. But traditional KV cache systems store each sequence separately — duplicating the exact same data over and over again. Pure waste.
March uses a Trie structure to automatically detect and reuse identical token prefixes. Instead of storing [system_prompt + history] 1000 times, it's stored once. Everyone shares it.
- 80-97% memory reduction in prefix-heavy workloads (tested on SmolLM2-135M with 500 multi-turn conversations)
- Zero-copy queries — returns direct pointers into the memory pool, no expensive memcpy on the hot path
- Predictable memory usage — fixed-size page pool with O(L) complexity
- Trade-off: slightly slower than dict O(1) lookup, but the memory savings are worth it in production posted an update 3 days ago
https://github.com/lizixi-0x2F/March
I just released March, an open-source high-performance KV cache sharing library for LLM inference that uses Trie-based prefix deduplication.
When you run LLM services, you often see thousands of requests sharing the same system prompt and conversation history. But traditional KV cache systems store each sequence separately — duplicating the exact same data over and over again. Pure waste.
March uses a Trie structure to automatically detect and reuse identical token prefixes. Instead of storing [system_prompt + history] 1000 times, it's stored once. Everyone shares it.
- 80-97% memory reduction in prefix-heavy workloads (tested on SmolLM2-135M with 500 multi-turn conversations)
- Zero-copy queries — returns direct pointers into the memory pool, no expensive memcpy on the hot path
- Predictable memory usage — fixed-size page pool with O(L) complexity
- Trade-off: slightly slower than dict O(1) lookup, but the memory savings are worth it in production reacted to reaperdoesntknow's post with 👍 5 days ago
We present a methodology for training small language models on CPU at FP32 precision
that achieves capability-per-dollar efficiency orders of magnitude beyond GPU-based training.
Across15modelsspanningfournovelarchitecturefamilies—MixtureofAttentions(MoA),cross-
architecture fusion (Qemma), swarm intelligence (SAGI), and metric-space causal language
models (DiscoverLM)—total compute cost was $24 on a single AMD EPYC 9454P proces-
sor. We introduce seven methodological pillars: (1) FP32 precision preservation, with exper-
iments demonstrating 5,810×single-operation error and 23,225×compounding error ratio for
FP16 at network depth; (2) sparse cognitive architectures where 0.02–7% of parameters activate
per token, matching CPU branching rather than GPU SIMD; (3) developmental curriculum
training progressing from language to logic to transfer to depth; (4) continuous belt-fed data
ingestion eliminating truncation waste; (5) hardware-native optimization for AMD Zen 4 via
AOCL/OpenMP/NUMA-aware allocation; (6) self-regulating thermodynamic governance with
emergent temperature measurement grounded in L2-star discrepancy; and (7) open-standard
compute (AVX2 SIMD at FP32) free of proprietary vendor dependency. We argue that transformers were designed for GPU hardware rather than mathematical optimality, and that architecture designed for geometric correctness—metric-space attention, triangle inequality enforcement, sparse expert routing—naturally favor CPU execution. For sub-2B parameter models, CPU training produces more capable models at a fraction of the cost.Organizations
