v.0.6.1-alpha

Install command, run in PowerShell. From an admin PowerShell it installs for the whole machine instead.

irm https://raw.githubusercontent.com/omgitsbase/llmash/main/install.ps1 | iex

On Linux, with sudo sh at the end to install for the whole machine:

curl -fsSL https://raw.githubusercontent.com/omgitsbase/llmash/main/install.sh | sh

Already installed: llmash update.

Fixed

• Qwen3.8 Flash Next keeps answering after a long prompt. Past about 200,000 tokens its replies turned to "!!!!" after the first token or two, and the slot kept failing until the model was reloaded. The fused convolution of its delta-net layers left the cells it saves for the next token unwritten, so each step started from stale memory. A 210,000-token prompt now continues the text, with the drafter keeping 45 to 63% of its drafts (0 before).
• A slot whose request went wrong recovers on the next request: a cleared recurrent state is now set to zero, where scaling it by zero kept a NaN.
• Flash Next's side slot reads the right cells: the two slots share one cache and each request reads all of it, since its sparse attention picks cells by their place in the whole cache.
• Many slots with a drafter no longer crash the server when requests arrive together right after a load (32 slots, a first wave of 2 or 8 requests on Gemma 4 E4B: every request in flight failed and the model reloaded). A slot lending itself to pad a round could be handed a new request while its draft was still on the card. Checked over seven such sequences from a cold load, none failing (each failed on 0.6.0).
• Many streams with a drafter (64 and more on Gemma 4 E4B) no longer stop the server when a slot without a request of its own is in the batch.
• A pinned model stays loaded once its keep-alive runs out, instead of showing "unloading" until it is unloaded by hand.
• llmash ablate shows its progress as the run goes, not in bursts.

Runtime

Build 789 (fork commit e70817ac1). Greedy replies of Gemma 4 E4B (8-bit, assistant drafter) and Qwen3.6 35B-A3B (NVFP4) are the ones build 782 gives, six prompts each.