Strata

fixtures

serve: DeepSeek-V4.1 through the server (deepseek-v3 pre-tokenizer, DSML tool calls, effort words, end token, card sampling)

2026-10-20 18:45:21 +02:00

web

Merge f-027-int2 into f-027-gh-app

2026-10-26 21:54:30 +01:00

__init__.py

Strata: Qwen3.8-Flash-Next on a 12-24 GB NVIDIA GPU + 64 GB RAM, one-click install

2026-09-24 18:42:20 +02:00

cache_drop.py

Start-up after another model: judge pinning by the emptiest NUMA node, say how much is file cache when RAM is short, opt-in STRATA_DROP_CACHE_ON_EXIT

2026-10-10 13:12:47 +02:00

chat_golden.json

Strata: Qwen3.8-Flash-Next on a 12-24 GB NVIDIA GPU + 64 GB RAM, one-click install

2026-09-24 18:42:20 +02:00

chat_template.jinja

Strata: Qwen3.8-Flash-Next on a 12-24 GB NVIDIA GPU + 64 GB RAM, one-click install

2026-09-24 18:42:20 +02:00

cpu_affinity.py

Group a launcher's allowed CPUs by physical core and split them between independent instances (serve/cpu_affinity.py, from #1436 )

2026-10-09 03:10:24 +02:00

ds_pictures.py

DeepSeek reads pictures by default: strata launch and setup pass --vision on (the engine's own 319 MiB tower, no separate reader), the server sends a picture to that tower as raw RGB in the file of a GENI line (serve/ds_pictures.py: decode with Pillow, both sides a multiple of 42, at most 1024 image tokens, the template's placeholder expanded to the span), --vision off keeps pictures off and the old 400; the plan, the catalog and docs/DEEPSEEK41.md say pictures on; tests for the size rule, the config switch, the prompt expansion and the server route

2026-10-27 05:41:48 +01:00

dsml.py

serve: a DeepSeek-V4.1 pack that ships the V4-Flash chat template (tool_calls, no space) is rendered with the V4.1 template, so tool calls come back as real calls; STRATA_DEBUG prints the raw ids; test

2026-10-24 00:07:13 +02:00

frontend.py

Parser: a fence is closed only by one as long as it opened with, and an indented code block is a quote, so calls shown in them are not run (#4702 #4703 )

2026-10-30 19:06:11 +01:00

images.py

Keep the card a person chose: strata launch no longer re-picks GPU 0, the engine's GPU 0 is explained, old hip_ordinal and list GPUs for images

2026-10-20 18:27:36 +02:00

keepawake.py

serve: opt-in prevent_sleep keeps Windows awake while a request is running or queued (#1727 )

2026-10-09 16:45:37 +02:00

launch_identity.py

serve: owned-server contract for the terminal (ready file, owner guard, launch identity, queue deadline, stage telemetry)

2026-10-15 10:23:23 +02:00

lifeline.py

serve: the lifeline identity is the child's start time, the sampler survives a bad source (#3590 , #3407 )

2026-10-26 18:15:01 +01:00

loader_diag.py

packaging: the Control python launcher uses no external command (dirname failed inside the AppImage on newer systems); the engine start error names a missing Windows DLL (0xC0000135) and says how to fix it

2026-10-19 10:52:51 +02:00

logit_bias.py

logit_bias: normalize a long list once per request instead of five times

2026-10-15 05:41:36 +02:00

mcp_fake_server.py

Tools from MCP servers in the web app's chat (serve/mcp.py)

2026-09-28 17:28:01 +02:00

mcp.py

MCP chat: opt-in mcp.image_results caches PNG/JPEG/WebP tool results beside the config, shows them in the web chat with an authenticated download and passes them to the vision reader (PR #2473 by cuijinkuo)

2026-10-26 18:22:31 +01:00

owner_guard.py

serve, control: a server's engine ends with it however it ends, leftovers are found by PID, start time and command line, runtime stop/cleanup recover a lost owner

2026-10-15 15:58:04 +02:00

prometheus.py

serve: GET /metrics in the Prometheus text format with vLLM's metric names (#793 )

2026-10-07 00:18:37 +02:00

ready.py

serve: owned-server contract for the terminal (ready file, owner guard, launch identity, queue deadline, stage telemetry)

2026-10-15 10:23:23 +02:00

replicas.py

shared arena on Windows: a named pagefile-backed section (CreateFileMapping/MapViewOfFile) with the same header and load state word, pid election on both OSes, server option enabled, pinned_shared_win_test

2026-10-30 06:02:30 +01:00

responses.py

serve: opt-in LaTeX-safe JSON recovery for structured output (#2893 )

2026-10-26 18:47:25 +01:00

runconfig.py

Windows file in use: every state-file swap goes through atomic.replace (retry); a sharing violation no longer says run as yourself

2026-10-20 14:04:57 +02:00

server.py

serve: the qualified GDN speedups are on from the server (Linux CUDA: GDN_TC + GDN_CONVL2=2; Linux HIP gfx12 only: GDN_CHUNKED), Qwen3.8 only, explicit env wins

2026-10-30 22:46:09 +01:00

session_save_retry.py

serve: wait for measured RAM before retrying reclaimed SAVE

2026-10-09 03:27:49 +02:00

split.py

Split serving, first version: GEN 0 prefill=1 in the engine, serve/split.py (SplitEngine, config split_serve / --split-serve), tests

2026-10-26 19:42:41 +01:00

structured.py

OpenAI logprobs, top_logprobs and allowed_token_ids on /v1/chat/completions (#3200 )

2026-10-26 19:08:35 +01:00

telemetry.py

serve: the lifeline identity is the child's start time, the sampler survives a bad source (#3590 , #3407 )

2026-10-26 18:15:01 +01:00

test_admit_order.py

serve/test_admit_order: the engine stand-in sets the arrivals deque the coalescing code reads

2026-10-30 18:39:58 +01:00

test_answer_speed.py

Monitor/serve: a one-token reply has no decode span, so no tok/s is shown for it (a 1-token probe read 1,250 tok/s); test

2026-10-25 09:29:40 +01:00

test_anthropic_base_url.py

Connect app: the third tab and strata connect give the server root (no /v1); the server also accepts /v1/v1/messages

2026-10-18 20:40:33 +02:00

test_cache_drop.py

Start-up after another model: judge pinning by the emptiest NUMA node, say how much is file cache when RAM is short, opt-in STRATA_DROP_CACHE_ON_EXIT

2026-10-10 13:12:47 +02:00

test_chat_tool_results.py

serve: Chat Completions tool results keep the order of the calls they answer (#3395 )

2026-10-26 18:54:34 +01:00

test_chunked_body.py

Limit body framing fix to incomplete chunked trailers

2026-10-09 03:04:21 +02:00

test_context_overflow.py

serve: shorten max_tokens to the room left instead of answering 400 (all three dialects)

2026-10-17 00:55:18 +02:00

test_control_cancel.py

serve: a Stop during the prompt read goes to the engine at once (#879 abort latency)

2026-10-05 11:03:50 +02:00

test_detok.py

tokenizer: _bpe reads self once instead of per symbol pair, and a threads-vs-serial encode test (#1385 ; the 'int has no attribute ranks' fault could not be reproduced and looks like an interpreter bug)

2026-10-07 17:21:48 +02:00

test_draft_layer.py

Models: a per-model draft layer (MTP) switch, picture-off and custom context that stick

2026-10-19 19:29:39 +02:00

test_ds_pictures.py

DeepSeek pictures: the Terminal's pictures step knows the tower is part of the engine (no reader lookup, no GPU/CPU reader choice), and a DeepSeek start without pictures answers a picture request with its own 400 wording

2026-10-27 06:12:25 +01:00

test_dsv41.py

Server: only a one-token end marker is a stop id; pieces of <|im_end|> no longer end a DeepSeek reply ("Nerazim" came out as "Neraz", #4576 )

2026-10-29 10:26:56 +01:00

test_effort_paths.py

Functional-test fixes: agent tok/s and context, plain tool rows, Thinking Off proof, profile used, no download wording for installed models, plain chat, smaller agent prompt

2026-10-16 04:20:20 +02:00

test_engine_arch_note.py

Server: one plain line at start when a graphics card in use is of an architecture the compiled engine was not built for (#4662 )

2026-10-30 19:18:45 +01:00

test_env_to_user.py

serve: "env_to_user" (opt-in) moves opencode's <env> block into the first user message, so a new folder or day reuses the cached system prompt (11.9 -> 1.1 s first request)

2026-10-13 00:48:16 +02:00

test_exif_orientation.py

vision: a JPEG with an EXIF Orientation other than 1 is turned upright before the encoder reads it (#1229 )

2026-10-07 03:02:56 +02:00

test_family_name.py

serve: DeepSeek-V4.1 through the server (deepseek-v3 pre-tokenizer, DSML tool calls, effort words, end token, card sampling)

2026-10-20 18:45:21 +02:00

test_fatal_recovery.py

serve: invalidate the engine after fatal verification timeouts

2026-10-05 13:16:53 +02:00

test_frontend.py

serve: a chat request with no message is a 400 (messages must be a non-empty list), as OpenAI answers it; test

2026-10-25 09:28:47 +01:00

test_gpu_roles_server.py

Server: gpu_roles (Main last, Off excluded, hand-edited gpu wins), roles log line, every-card-Off handling

2026-10-20 18:36:30 +02:00

test_history_tools.py

Server: a tool the conversation already called counts as declared, so a call to it comes back as a tool call, not raw text (#4663 )

2026-10-30 19:06:50 +01:00

test_images.py

Keep the card a person chose: strata launch no longer re-picks GPU 0, the engine's GPU 0 is explained, old hip_ordinal and list GPUs for images

2026-10-20 18:27:36 +02:00

test_keepawake.py

serve: opt-in prevent_sleep keeps Windows awake while a request is running or queued (#1727 )

2026-10-09 16:45:37 +02:00

test_leftover_engines.py

serve: an engine left behind by an earlier run is waited for until it has given its memory back (#4051 )

2026-10-29 00:01:54 +01:00

test_lifecycle.py

serve, control: a server's engine ends with it however it ends, leftovers are found by PID, start time and command line, runtime stop/cleanup recover a lost owner

2026-10-15 15:58:04 +02:00

test_lifeline.py

serve: the lifeline identity is the child's start time, the sampler survives a bad source (#3590 , #3407 )

2026-10-26 18:15:01 +01:00

test_logit_bias.py

logit_bias: normalize a long list once per request instead of five times

2026-10-15 05:41:36 +02:00

test_logprobs_parallel.py

OpenAI logprobs, top_logprobs and allowed_token_ids on /v1/chat/completions (#3200 )

2026-10-26 19:08:35 +01:00

test_logprobs.py

OpenAI logprobs, top_logprobs and allowed_token_ids on /v1/chat/completions (#3200 )

2026-10-26 19:08:35 +01:00

test_mcp_media.py

MCP chat: opt-in mcp.image_results caches PNG/JPEG/WebP tool results beside the config, shows them in the web chat with an authenticated download and passes them to the vision reader (PR #2473 by cuijinkuo)

2026-10-26 18:22:31 +01:00

test_mcp.py

MCP chat: at the tool-round limit the model gets one final reply with no tools from the results already found, instead of the turn ending without an answer (PR #3741 by i-Vladimir-i)

2026-10-26 18:19:27 +01:00

test_model_card_sampling.py

serve: a request with no sampling gets the model card's (thinking 1.0/0.95/20; not thinking 0.7/0.8/20 + presence 1.5; Qwen3.6 too); greedy only when asked (#2685 )

2026-10-15 22:35:55 +02:00

test_monitor.py

serve: shorten max_tokens to the room left instead of answering 400 (all three dialects)

2026-10-17 00:55:18 +02:00

test_native_tool_parser.py

serve: tool-call aliases and wrapper boundaries (port of PR #3556 ), and an unparsed call is logged (T1)

2026-10-22 13:01:04 +02:00

test_parallel_yield.py

Fix #2289 : Anthropic usage reports cache_creation_input_tokens, the prompt tokens read now and kept for the next request (from PR #2290 )

2026-10-20 14:13:34 +02:00

test_parallel.py

serve: the qualified GDN speedups are on from the server (Linux CUDA: GDN_TC + GDN_CONVL2=2; Linux HIP gfx12 only: GDN_CHUNKED), Qwen3.8 only, explicit env wins

2026-10-30 22:46:09 +01:00

test_parser_call_scan.py

serve: skip tool-body rescans until a closing tag arrives

2026-10-14 12:49:34 +02:00

test_parser_hold.py

serve: avoid exhaustive partial-tag suffix checks

2026-10-14 12:49:26 +02:00

test_parser_markdown_code.py

Parser: a fence is closed only by one as long as it opened with, and an indented code block is a quote, so calls shown in them are not run (#4702 #4703 )

2026-10-30 19:06:11 +01:00

test_parser_track.py

Parser: a fence is closed only by one as long as it opened with, and an indented code block is a quote, so calls shown in them are not run (#4702 #4703 )

2026-10-30 19:06:11 +01:00

test_pool_cores.py

serve, launch: "pool_cores" in the run config and strata launch --pool-cores

2026-10-20 19:16:17 +02:00

test_prometheus.py

serve: GET /metrics in the Prometheus text format with vLLM's metric names (#793 )

2026-10-07 00:18:37 +02:00

test_prompt_encoder.py

serve tests: the refused-prompt test runs in strict mode (fit_max_tokens off) so a huge max_tokens is refused, not shortened

2026-10-17 07:00:49 +02:00

test_q36_config.py

Qwen3.6-35B-A3B: setup asks for pictures and recommends parallel requests; docs

2026-10-14 01:11:11 +02:00

test_queue_keepalive.py

serve: a request waiting for the single turn sends keep-alives, and gives its place up when cancelled (#1619 )

2026-10-09 03:15:12 +02:00

test_quiet_cancel.py

serve: poll the cancel event every 0.5 s while the engine is quiet, so a client that leaves during a long prompt read stops it (#1182 )

2026-10-07 03:02:56 +02:00

test_quoted_call_values.py

serve: only a whole quoted <function=NAME> opener keeps a parameter value open (#4078 )

2026-10-28 23:33:10 +01:00

test_reader_follows_engine.py

Server: a picture reader written into an older engine folder follows the engine in use after an update (#4658 )

2026-10-30 19:03:21 +01:00

test_ready_timeout.py

Fix #2496 : an engine program that cannot be started is a 503 with the reason, not a dropped connection (from PR #2497 )

2026-10-20 14:13:34 +02:00

test_reasoning_loop_recovery.py

serve: max_tokens counts the thinking - without a budget, thinking is closed early enough to leave a quarter of the cap for the answer (cap 8192 cut the thinking off with no answer); config answer_room:false opts out; tests

2026-10-16 00:38:04 +02:00

test_reasoning_rescue.py

serve: a tool call inside a code fence or inline code in the visible answer is text (#1058 )

2026-10-06 11:25:45 +02:00

test_reasoning_tools.py

serve: a </think> the reasoning quotes stays reasoning instead of ending it (#537 )

2026-10-09 03:15:12 +02:00

test_replicas.py

shared arena on Windows: a named pagefile-backed section (CreateFileMapping/MapViewOfFile) with the same header and load state word, pid election on both OSes, server option enabled, pinned_shared_win_test

2026-10-30 06:02:30 +01:00

test_request_accounting.py

Fix #2289 : Anthropic usage reports cache_creation_input_tokens, the prompt tokens read now and kept for the next request (from PR #2290 )

2026-10-20 14:13:34 +02:00

test_request_hardening.py

serve: repeat_penalty / repeat_last_n (llama.cpp's names) for the config's sampling block and for requests (#1819 )

2026-10-10 10:45:03 +02:00

test_responses.py

serve: /v1/responses accepts max_output_tokens of 0 and -1 (#4036 )

2026-10-26 18:45:04 +01:00

test_restart_waiters.py

serve: a request the engine refuses with ERR answers at once, not after a 300 s wait for a DONE (#1059 )

2026-10-06 15:57:37 +02:00

test_runconfig.py

web: a Model settings view for documented run config keys (#564 )

2026-10-03 18:35:02 +02:00

test_security.py

serve: support IPv6 bind addresses (#3212 )

2026-10-26 18:50:11 +01:00

test_server.py

serve: tool_call_recovery is on by default; "tool_call_recovery": false is the off switch

2026-10-29 10:25:07 +01:00

test_session_save_retry.py

serve: wait for measured RAM before retrying reclaimed SAVE

2026-10-09 03:27:49 +02:00

test_slot_alloc.py

serve: pick_slot sends a chat's next turn to the slot that holds it, found by its turn checkpoint (#2518 )

2026-10-15 05:59:33 +02:00

test_slots.py

test_slots: compare the relative save dir through realpath (8.3 temp dir names on Windows)

2026-10-09 14:47:49 +02:00

test_split.py

Split serving, first version: GEN 0 prefill=1 in the engine, serve/split.py (SplitEngine, config split_serve / --split-serve), tests

2026-10-26 19:42:41 +01:00

test_stop_holdback.py

Server: only a one-token end marker is a stop id; pieces of <|im_end|> no longer end a DeepSeek reply ("Nerazim" came out as "Neraz", #4576 )

2026-10-29 10:26:56 +01:00

test_structured.py

serve: opt-in LaTeX-safe JSON recovery for structured output (#2893 )

2026-10-26 18:47:25 +01:00

test_tool_call_envelope.py

serve: tool-call aliases and wrapper boundaries (port of PR #3556 ), and an unparsed call is logged (T1)

2026-10-22 13:01:04 +02:00

test_tool_call_unclosed.py

serve: tool-call aliases and wrapper boundaries (port of PR #3556 ), and an unparsed call is logged (T1)

2026-10-22 13:01:04 +02:00

test_tool_corpus.py

serve: damaged tool-call arguments, a stray closing think tag and a missing </function> no longer cost the call (#4427 #4077 )

2026-10-28 23:55:33 +01:00

test_tool_name_alias.py

serve: a call returned as text for an undeclared name is named in the reply; the collapsed mcp__a__b spelling is the declared tool (#3796 )

2026-10-26 20:30:44 +01:00

test_tool_name_space.py

serve: a tool call written <function= NAME> is the call NAME

2026-10-08 02:25:52 +02:00

test_tool_routes.py

serve: damaged tool-call arguments, a stray closing think tag and a missing </function> no longer cost the call (#4427 #4077 )

2026-10-28 23:55:33 +01:00

test_vision_busy.py

serve, control: a server's engine ends with it however it ends, leftovers are found by PID, start time and command line, runtime stop/cleanup recover a lost owner

2026-10-15 15:58:04 +02:00

test_vision_elastic_http.py

serve: return 503 when the elastic image encoder cannot start (#2435 )

2026-10-26 18:51:44 +01:00

test_vision_elastic.py

serve: return 503 when the elastic image encoder cannot start (#2435 )

2026-10-26 18:51:44 +01:00

test_vision_start_error.py

serve: a tool's picture with an elastic image encoder starts the encoder (I6); a failed encoder start carries its stderr tail, exit status and card hint (I3)

2026-10-22 12:52:37 +02:00

test_vision_token_limit.py

vision: the image-token limit has a default and is enforced (I7, #3690 )

2026-10-22 13:10:12 +02:00

test_vision_tool_image.py

serve: a tool's picture with an elastic image encoder starts the encoder (I6); a failed encoder start carries its stderr tail, exit status and card hint (I3)

2026-10-22 12:52:37 +02:00

test_vision_workdir.py

Fix #3370 : the vision work folder is made again when a temp cleaner removed it

2026-10-20 14:13:34 +02:00

test_vram.py

Hot VRAM resize, opt-in: a segmented expert cache, VRAM <reserve>, POST /v1/vram (#533 )

2026-10-03 18:52:00 +02:00

test_web_api_key.py

Web page: use an edited API key at once, even when browser storage is blocked (#3435 )

2026-10-22 17:32:35 +02:00

test_web_badges.py

Web page: a request that timed out in the queue shows a Timeout badge in Recent requests (PR #3991 by thisisspartan)

2026-10-26 18:17:49 +01:00

test_web_follow.py

Chat: the transcript follows the answer only while the reader has not left; a wheel notch, finger or key moving up lets go at once, in the app chat and the served web page (#4014 )

2026-10-26 18:11:11 +01:00

test_web_history.js

Web chat: one chat-history design - saved chats, edit and rewind, branch, send while answering (#1787 , #2615 , #3823 , #4004 )

2026-10-26 20:14:14 +01:00

test_web_history.py

Web chat: one chat-history design - saved chats, edit and rewind, branch, send while answering (#1787 , #2615 , #3823 , #4004 )

2026-10-26 20:14:14 +01:00

test_web_i18n.js

Web page: a translation mechanism, and a language choice in About > Settings

2026-10-26 19:52:26 +01:00

test_web_i18n.py

Web page: a translation mechanism, and a language choice in About > Settings

2026-10-26 19:52:26 +01:00

test_web_load_unload.py

Web page: Load and Unload in the header (#3550 )

2026-10-26 19:59:54 +01:00

test_web_math.js

Web chat: draw math with the bundled KaTeX, opt-in and offline (#2972 , #2974 )

2026-10-26 19:52:56 +01:00

test_web_math.py

Web chat: draw math with the bundled KaTeX, opt-in and offline (#2972 , #2974 )

2026-10-26 19:52:56 +01:00

test_web_preview.js

Web chat: preview a generated svg or html block in a sandboxed frame (#1628 )

2026-10-26 19:57:56 +01:00

test_web_preview.py

Web chat: preview a generated svg or html block in a sandboxed frame (#1628 )

2026-10-26 19:57:56 +01:00

test_web_sampling.py

web: the page sends no sampling until the user changes it, so the server's model-card defaults apply (stored 0.6 counts as not set); test (#2685 )

2026-10-15 23:34:15 +02:00

test_winjob.py

winjob: opt the server and its children out of Windows power throttling (#691 )

2026-10-05 15:58:36 +02:00

webfiles.py

Web page: a translation mechanism, and a language choice in About > Settings

2026-10-26 19:52:26 +01:00

winjob.py

runtime: the managed owner runtime on Windows (loopback TCP + token, job object, byte-range leases)

2026-10-15 12:25:58 +02:00