标签：#attention

DeepSeek has previewed its V4 model family, led by a 1.6 trillion-parameter flagship, and made a one-million-token context window the default across all its services. The weights are downloadable and self-hostable, putting frontier-scale long context in reach of smaller labs and individuals without per-token payment to a closed provider. Key facts What: DeepSeek previewed two free-to-download V4 models that can read a million tokens at once, no longer as a premium add-on but as the standard setting. When: 2026-06-29 Primary source: read the source A large language model has no persistent memory. Each time it answers, it re-reads everything in front of it — your question, the conversation so far, any documents you pasted — and that pile of text is the context. The context window is the hard ceiling on how much it can hold at once. For years that ceiling was a few thousand words, then tens of thousands. Pushing it to a million has been possible but expensive, usually sold as a special, pricey tier. DeepSeek's move is to make a million the everyday default. The family comes in two sizes. V4-Pro is the big one — 1.6 trillion parameters in total, but only about 49 billion of them switch on for any given word. That design is called a mixture of experts : instead of running the entire brain for every token, the model routes each piece of text to a small relevant subset of specialists, so it stays affordable to run despite its enormous size. V4-Flash is the smaller, cheaper, faster sibling, meant for everyday chat and quick edits, and DeepSeek says it keeps up with Pro on simpler agent tasks. Making a million-token window affordable comes down to how the model handles its KV cache — the running set of notes it stores about every previous word, which grows steadily the longer the conversation gets. At a million tokens those notes become a mountain of memory, and the model normally has to consult every note for every new word it writes. DeepSeek's approach, which they call sp

What flipped in b9437 Build b9437 , published on May 30, 2026 at 20:56 UTC , ships two targeted default-value corrections to llama-bench . Flash attention ( -fa ) shifts from a hard-coded off to auto ( LLAMA_FLASH_ATTN_TYPE_AUTO ), and the GPU-layer count ( -ngl ) changes from the legacy sentinel 99 to -1 . Both values now match what llama-server and llama-cli already used — the bench tool was simply never updated to track them until this build. Quick Answer: Before b9437 (published May 30, 2026) , llama-bench hard-coded -fa off , silently skipping flash attention even on CUDA, Metal, and Vulkan hardware. Build b9437 sets the default to -fa auto and -ngl -1 , matching llama-server and llama-cli . Any pre-b9437 baseline on FA-capable hardware needs a flag-matched re-run to remain valid. PR #23714 , reviewed and merged by maintainers JohannesGaessler and pwilkin, adds the same -fa auto|off|on tri-state flag to llama-bench that the rest of the toolchain already supported. With LLAMA_FLASH_ATTN_TYPE_AUTO as the new default, flash attention activates automatically when the runtime detects a capable backend (CUDA, Metal, Vulkan); on CPU-only hosts it stays off with no error and no output change. Parameter Before b9437 After b9437 Behavioral impact -fa off (hard-coded) auto ( LLAMA_FLASH_ATTN_TYPE_AUTO ) GPU-capable hosts bench with FA active by default; pre/post comparisons require explicit flag-matching -ngl 99 (offload-all sentinel) -1 (runtime decides) CPU-only builds no longer attempt full GPU offload; eliminates spurious CUDA errors when no GPU is present The following verified script (executed successfully, exit 0) demonstrates the behavioral gap in concrete terms — on a capable GPU, the pre-b9437 defaults schedule zero FA rows while b9437 defaults schedule one: def old_llama_bench ( device ): # Before b9437, the default bench matrix used FA=0, so FA rows were skipped. return [{ " device " : device [ " name " ], " ngl " : 0 , " fa " : 0 }] def b9437_llama_bench ( de