Skip to content

Firetitan server - #4

Draft
1stprinciple wants to merge 7 commits into
mainfrom
firetitan-server
Draft

Firetitan server#4
1stprinciple wants to merge 7 commits into
mainfrom
firetitan-server

Conversation

@1stprinciple

Copy link
Copy Markdown
Owner

No description provided.

@1stprinciple
1stprinciple force-pushed the firetitan-server branch 3 times, most recently from 0b67bfc to 671c8fc Compare March 30, 2026 22:21
@1stprinciple
1stprinciple force-pushed the firetitan-server branch 2 times, most recently from 48fa4ec to debad83 Compare April 7, 2026 02:12
http_timeout=120.0

cursor add retry

hack sampler_path

checkpoint_type="base"

comment out kl_sample_train_metrics

max_concurrency = 8

default 4

cursor added max_concurrency

print

sampling_client = DeploymentSampler( for async

0

['choices']

contents

logprobs=True,

print("sample_result:", sample_result)

remove max_seq_len

max_tokens=19200

field

add back dataclass

FireworksTokenCompleter in metric_util

fix tokenizer

http://localhost:8902

remove dataclass for FireworksTokenCompleter

state_name, sampler_name

FireworksTokenCompleter

eval_semaphore

replace firetitan in rl path

squash

dash

comment compute_mean_nll

print fwd_bwd_result

print(f"fwd_bwd_result.loss_fn_outputs

print(f"len(logprobs) = {len(logprobs)}")

base_url: str = "http://localhost:8099"

import FiretitanTrainingClient

no await on service_client.create_training_client

lora_rank=config.lora_rank

fix FiretitanTrainingClient import

Replace for SFT

Revert "train_unembed=False"

This reverts commit 89c976d.

Revert "rank"

This reverts commit e14eeb4.

Revert "use tinker client temporarily"

This reverts commit 0106c20.

Revert "create_lora_training_client"

This reverts commit e905aa6.

rank

create_lora_training_client

use tinker client temporarily

train_unembed=False

Revert "attach grads as well"

This reverts commit 5ab4387.

attach grads as well

remove last_call_logprobs

base_url = None

write weights as well.

Claude hack

Add custom loss

add log

resume_info

open-research

train_on_what = TrainOnWhat.LAST_ASSISTANT_MESSAGE

unmatch )

Support Qwen3-30B-A3B-Thinking-2507

revert some changes.

delta checkpoint

max_concurrency: int = 2

max_concurrency: int = 4 for 1 GPU

use FireworksTokenCompleter in rollouts.py

path_dict

fireworks_base_model_name

save initially

Revert "try cursor fix (i don't think it's correct)"

This reverts commit b772b9f.

try cursor fix (i don't think it's correct)

sync

Add fireworks_deployment_id

remove api_key in deploymentsampler

fireworks_hot_load_timeout

ordering weight_syncer

add weight_syncer

now setup runs

print with hydra style

remove reference of self.

copy over

ppo

sample_result, _

async_completions_stream

use built-in grpo

use ppo_loss_fn instead

ALL_ASSISTANT_MESSAGES for open researcher

policy_ep

Add custom RL loss

works

use lora

fireworks_deployment_id in CLIConfig

cfg

revert

Cursor fix test_group_size of twenty questions

support more multiplay_rl

fireworks specific for multiplayer_rl

loss-fn

sample_kwargs

_training_logprobs_from_forward

tokenizer

27b

remove stop condition

print

try catch

stop=stop

checkpoint_type = "base"

Qwen3.5-27B

kimi thinking_parts

initial_messages

remove fwd_future

deep-research-sft-0406

fireworks_base_model_name

Add distillation

no dcp_timeout

if weight_syncer is not None and weight_syncer.base_identity is not None:
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant