github-starred/ollama
2
0
Fork 0
mirror of https://github.com/ollama/ollama.git synced 2026-05-06 16:11:34 -05:00
Code Issues 14.5k Packages Projects Releases 120 Wiki Activity

[GH-ISSUE #12443] Different load times for an already loaded model, Why? #8266

New Issue
Open
opened 2026-04-12 20:48:26 -05:00 by GiteaMirror · 0 comments
GiteaMirror commented 2026-04-12 20:48:26 -05:00
Owner
Copy Link

Originally created by @yisuscc on GitHub (Sep 29, 2025).
Original GitHub issue: https://github.com/ollama/ollama/issues/12443

I'm currently doing a benchmarking tool with ollama and noticed that it reports a different load time for each prompt fed (see the column load_duration in the picture below) when the model should stay loaded in memory (the keep_alive is set at -1 ). The same thing happens in all models tested (gemma3n, mistral,qwen2.5:0.5b ...)

Image

For comparison, here is the llama.cpp results for a different model, notice how the load duration is the same across all prompts.

Image

The inference code for both versions:

  • Ollama:

def query_event(self, prompt: str, model: str, event: Event, prompt_id: int = -1, keep_alive='-1') -> PromptMetrics: start = time.time_ns() event.set() response: GenerateResponse = self.client.generate(prompt=prompt, model=model, keep_alive=keep_alive) finish = time.time_ns() event.clear() return PromptMetrics.ollama_pseudoconstructor(start, finish, response, prompt_id)

  • llama.cpp:
    def query_event(self,prompt:str,event:Event, prompt_id=-1)-> PromptMetrics: #limpiaos el contexto de rendimiento llama_perf_context_reset(self.ctx) event.set() #so the time is a bit more accurate start= time.time_ns() #We query the machine response= self(prompt) finish = time.time_ns() event.clear() return PromptMetrics.llama_cpp_pseudoconstructor(start, finish, self.get_performance_metrics(), self.get_name(), prompt_id)
Originally created by @yisuscc on GitHub (Sep 29, 2025). Original GitHub issue: https://github.com/ollama/ollama/issues/12443 I'm currently doing a benchmarking tool with ollama and noticed that it reports a different load time for each prompt fed (see the column _load_duration_ in the picture below) when the model should stay loaded in memory (the keep_alive is set at -1 ). The same thing happens in all models tested (gemma3n, mistral,qwen2.5:0.5b ...) <img width="1421" height="107" alt="Image" src="https://github.com/user-attachments/assets/329b63ed-0b33-4ad6-8440-537db10f24ec" /> For comparison, here is the llama.cpp results for a different model, notice how the load duration is the same across all prompts. <img width="1421" height="107" alt="Image" src="https://github.com/user-attachments/assets/4831a2fd-1924-4d52-9744-b6940617d8ed" /> The inference code for both versions: - Ollama: ` def query_event(self, prompt: str, model: str, event: Event, prompt_id: int = -1, keep_alive='-1') -> PromptMetrics: start = time.time_ns() event.set() response: GenerateResponse = self.client.generate(prompt=prompt, model=model, keep_alive=keep_alive) finish = time.time_ns() event.clear() return PromptMetrics.ollama_pseudoconstructor(start, finish, response, prompt_id)` - llama.cpp: `def query_event(self,prompt:str,event:Event, prompt_id=-1)-> PromptMetrics: #limpiaos el contexto de rendimiento llama_perf_context_reset(self.ctx) event.set() #so the time is a bit more accurate start= time.time_ns() #We query the machine response= self(prompt) finish = time.time_ns() event.clear() return PromptMetrics.llama_cpp_pseudoconstructor(start, finish, self.get_performance_metrics(), self.get_name(), prompt_id)`
Sign in to join this conversation.
No Branch/Tag Specified
Branches Tags
Labels
Clear labels
amd
api
app
bug
build
cli
cloud
compatibility
context-length
create
docker
documentation
embeddings
feature request
feedback wanted
good first issue
gpt-oss
gpu
harmony
help wanted
image
install
intel
js
launch
linux
macos
memory
mlx
model
needs more info
networking
nvidia
ollama.com
performance
pull-request
Mirrored from GitHub Pull Request
 python
question
registry
rendering
thinking
tools
top
vulkan
windows
wsl
No Label
Milestone
No items
No Milestone
Projects
Clear projects
No project
Assignees
Clear assignees
GiteaMirror ninjasurge
No Assignees
1 Participants
Notifications
Due Date
No due date set.
Dependencies

No dependencies set.

Reference: github-starred/ollama#8266
Delete Branch "%!s()"

Deleting a branch is permanent. Although the deleted branch may continue to exist for a short time before it actually gets removed, it CANNOT be undone in most cases. Continue?