python-fastllm 0.0.61__py3-none-any.whl → 0.0.63__py3-none-any.whl

fastllm/__init__.py

@@ -2,4 +2,4 @@
 
 - `fastllm.responses`: Provider-independent Responses resources and streaming over FastLLM"""
 
-__version__ = "0.0.61"
+__version__ = "0.0.63"

fastllm/acomplete.py

@@ -168,20 +168,23 @@ async def acomplete(msgs, model, api_name=None, vendor_name=None, api_key=None,
     if own:
         vendor_name = ifnone(vendor_name, api_name)  # an own-transport api: no HTTP client to build
         if oauth_token: kwargs['oauth_token'] = oauth_token  # its `mk_payload` applies the user's token itself
-    else: cli, api_name, vendor_name = mk_client(model=model, vendor_name=vendor_name, api_name=api_name, api_key=api_key, oauth_token=oauth_token, base_url=base_url, xtra_hdrs=xtra_hdrs)
+    else: cli, api_name, vendor_name = mk_client(model=model, vendor_name=vendor_name, api_name=api_name, api_key=api_key, oauth_token=oauth_token, base_url=base_url)
     api = api_registry[api_name]
     if previous_response_id is not None:
         if not getattr(api, 'supports_previous_response_id', False):
             raise ValueError('previous_response_id is not supported by this transport')
         kwargs['previous_response_id'] = previous_response_id
     payload = api.mk_payload(msgs, model, stream=stream, **kwargs)
-    xtra_body = ifnone(xtra_body, {})
     if fix := getattr(api, 'fix_payload', None): fix(payload, model, vendor_name)
+    body = payload | ifnone(xtra_body, {})
     if not own:
+        headers = xtra_hdrs or {}
+        if vendor_name == 'codex' and (key := body.get('prompt_cache_key')):
+            if not any(k.lower() == 'session-id' for k in headers): headers = {**headers, 'session-id':key}
         ep = endpoint or api.endpoint
         path,params = ep(model, stream) if callable(ep) else (ep, None)
-    if defaults.debug_mode: _debug_print(model, api_name, vendor_name, payload | xtra_body, None if own else cli._url(path))
-    async def _call(): return await provider_req(cli, path, payload | xtra_body, params=params, stream=stream)
+    if defaults.debug_mode: _debug_print(model, api_name, vendor_name, body, None if own else cli._url(path))
+    async def _call(): return await provider_req(cli, path, body, params=params, stream=stream, headers=headers)
     async def _mk_gen():
         resp = payload if own else await _call()  # an own-transport api's `acollect_stream` makes the request itself, from the payload
         async for o in api.acollect_stream(resp, model=model, vendor_name=vendor_name, stop_callables=stop_callables): yield o

fastllm/chat.py

@@ -9,6 +9,7 @@ __all__ = ['effort', 'contents', 'stop_reason', 'cite_footnote', 'postproc', 'Re
            'StopSequencesCallback', 'StreamAccum', 'adisplay_stream']
 
 # %% ../nbs/07_chat.ipynb #d5a3bc1f
+import secrets
 from typing import Optional,Callable
 from fastcore.funccall import mk_ns, call_func, call_func_async, get_schema
 from fastcore.utils import *
@@ -168,6 +169,7 @@ class AsyncChat:
         cache=False,              # Anthropic prompt caching
         cache_idxs:list=[-1],     # Anthropic cache breakpoint idxs, use `0` for sys prompt if provided
         ttl=None,                 # Anthropic prompt caching ttl
+        prompt_cache_key=None,     # Stable conversation key for OpenAI Responses; generated once if omitted
         api_name=None,            # API to use, one of ApiName: openai (responses), openai_chat, anthropic, gemini
         vendor_name=None,         # Vendor name, one of vendor_mapping which resolves api_base/api_key automatically
         api_key=None,             # API key when model can't be resolved or vendor_name is not known or codex
@@ -195,6 +197,7 @@ class AsyncChat:
         self.response_id,self._response_hist_idx = None,0
         self._turn_start = 0    # index into `hist` where the current turn began, for `full`
         self.last_req_use = None  # usage of the latest request only; `use` accumulates across a turn's tool-call steps
+        if prompt_cache_key is None: prompt_cache_key = secrets.token_hex(16)
         store_attr(but='cbs')
         self.cbs = L()
         if default_cbs: self.add_cbs(defaults.chat_callbacks)
@@ -256,6 +259,7 @@ def _prep_call(self:AsyncChat, search, max_tokens, kwargs, stream=False, think=N
     if self.base_url:      kwargs['base_url'] = self.base_url
     if self.endpoint:      kwargs['endpoint'] = self.endpoint
     if self.extra_headers: kwargs['xtra_hdrs'] = self.extra_headers
+    kwargs.setdefault('prompt_cache_key', self.prompt_cache_key)
     kwargs.update(_think_kw(self.model, think, self.vendor_name))
     return max_tokens
 

fastllm/openai_responses.py

@@ -257,6 +257,7 @@ def mk_payload(msgs, model, **kwargs):
     if tools:=kwargs.get('tools'):          payload['tools'] = denorm_tool_schs(tools)
     if tchc:=kwargs.get('tool_choice'):     payload['tool_choice'] = denorm_tool_choice(tchc)
     if rid:=kwargs.get('previous_response_id'): payload['previous_response_id'] = rid
+    if key:=kwargs.get('prompt_cache_key'): payload['prompt_cache_key'] = key
     if (ptc:=kwargs.get('parallel_tool_calls')) is not None: payload['parallel_tool_calls'] = ptc
     if thk:=kwargs.get('reasoning_effort'): payload['reasoning'] = denorm_reasoning(thk)
     if (wopts:=kwargs.get('web_search_options')) is not None:  payload.setdefault('tools', []).append(denorm_web_search(wopts))

fastllm/responses.py

@@ -4,7 +4,7 @@
 
 `normalize_call_ids` supplies missing client call IDs before we retain the history. Later tool results can refer to the same IDs that the client received.
 
-`ResponseState` holds the public ID, model, full message history, and optional provider response ID. The host can store it under its own ownership and expiry rules. A provider ID lets the next call send just the new messages; without one, it needs the retained history.
+`ResponseState` holds the public ID, model, full message history, prompt cache key, and optional provider response ID. The host can store it under its own ownership and expiry rules. A provider ID lets the next call send just the new messages; without one, it needs the retained history.
 
 FastLLM calls stream internally. The helpers here turn their parts into Responses SSE frames; they don't decide storage, ownership, or billing policy."""
 
@@ -133,7 +133,7 @@ def response_usage(usage):
 def _response_base(rid, body, previous_response_id=None, created_at=None):
     result = dict(id=rid, object='response', created_at=int(created_at or time.time()), instructions=body.get('instructions'), model=body['model'])
     result.update(parallel_tool_calls=body.get('parallel_tool_calls', True), previous_response_id=previous_response_id,
-        reasoning=body.get('reasoning'), store=True, temperature=body.get('temperature'))
+        reasoning=body.get('reasoning'), store=True, prompt_cache_key=body.get('prompt_cache_key'), temperature=body.get('temperature'))
     result.update(text=body.get('text', {'format':{'type':'text'}}), tool_choice=body.get('tool_choice', 'auto'),
         tools=body.get('tools', []), top_p=body.get('top_p'), metadata=body.get('metadata', {}))
     return result
@@ -152,18 +152,19 @@ def response_object(rid, body, comp, previous_response_id=None, created_at=None)
 # %% ../nbs/06a_responses.ipynb #73a2fc16
 class ResponseState(BasicRepr):
     "Canonical history and optional provider continuation metadata."
-    def __init__(self, id, model, history, provider_response_id=None): store_attr()
+    def __init__(self, id, model, history, provider_response_id=None, prompt_cache_key=None): store_attr()
 
 # %% ../nbs/06a_responses.ipynb #b28fa30d
 @patch
 def to_dict(self:ResponseState):
     "This state as a JSON-ready dict"
-    return dict(id=self.id, model=self.model, history=[msg2dict(m) for m in self.history], provider_response_id=self.provider_response_id)
+    return dict(id=self.id, model=self.model, history=[msg2dict(m) for m in self.history],
+        provider_response_id=self.provider_response_id, prompt_cache_key=self.prompt_cache_key)
 
 @patch(cls_method=True)
 def from_dict(cls:ResponseState, d):
     "The state `to_dict` produced `d` from"
-    return cls(d['id'], d['model'], tuple(dict2msg(m) for m in d['history']), d['provider_response_id'])
+    return cls(d['id'], d['model'], tuple(dict2msg(m) for m in d['history']), d['provider_response_id'], d['prompt_cache_key'])
 
 # %% ../nbs/06a_responses.ipynb #74e09d18
 class ResponseStore:
@@ -221,6 +222,8 @@ class AsyncResponses:
                 if isinstance(p, ToolResult) and not p.name: p.name = names.get(p.id, '')
         instructions = '\n\n'.join(filter(None, [body.get('instructions', ''), input_system]))
         provider_messages = new_msgs if previous and previous.provider_response_id else list(history)
+        key = body.get('prompt_cache_key') or (previous.prompt_cache_key if previous else None) or secrets.token_hex(16)
+        body = body | dict(prompt_cache_key=key)
         return ResponseTurn(_new_id('resp'), int(time.time()), body, history, provider_messages, instructions, previous,
             {f.__name__: f for f in server_tools or []}, kwargs)
 
@@ -233,7 +236,7 @@ async def call(self:AsyncResponses, turn):
     reasoning = body.get('reasoning') or {}
     tools = list(body.get('tools') or []) + [dict(type='function', **get_schema(f, pname='parameters')) for f in turn.server_tools.values()]
     kwargs = dict(previous_response_id=turn.provider_previous_id, system=turn.system or None, tools=tools or None)
-    kwargs.update(tool_choice=body.get('tool_choice'), parallel_tool_calls=body.get('parallel_tool_calls', True),
+    kwargs.update(prompt_cache_key=body['prompt_cache_key'], tool_choice=body.get('tool_choice'), parallel_tool_calls=body.get('parallel_tool_calls', True),
         reasoning_effort=reasoning.get('effort'), max_tokens=body.get('max_output_tokens'), temperature=body.get('temperature'),
         cache_idxs=body.get('cache_idxs'), ttl=body.get('ttl'), web_search_options=body.get('web_search_options'))
     stream = await acomplete(turn.provider_messages, model=body['model'], stream=True, **kwargs, **turn.call_kwargs)
@@ -250,7 +253,7 @@ def state(self:AsyncResponses, turn, comp, results=()):
     "Build the next continuation state; `results` are server tool results answering `comp`'s calls."
     normalize_call_ids(comp)
     history = (*turn.history, comp.message, *([Msg('tool', list(results))] if results else []))
-    return ResponseState(turn.id, turn.body['model'], history, comp.response_id)
+    return ResponseState(turn.id, turn.body['model'], history, comp.response_id, turn.body['prompt_cache_key'])
 
 
 # %% ../nbs/06a_responses.ipynb #7615b274

fastllm/types.py

@@ -4,12 +4,12 @@
 
 # %% auto #0
 __all__ = ['FinishReason', 'api_registry', 'model_prices_url', 'sample_img_url', 'sample_doms', 'haik45', 'sonn45', 'sonn46',
-           'sonn', 'sonn5', 'opus46', 'opus48', 'opus', 'opus5', 'fable', 'fable5', 'fable51', 'gpt54', 'gpt54m',
-           'gpt55', 'codex54', 'codex54m', 'codex55', 'codex53spark', 'model_info_registry', 'modern_llm',
-           'effort_codes', 'mimo_v25_common', 'codex_pricing', 'sol', 'terra', 'luna', 'gpt56s', 'astra', 'Usage',
-           'APIRegistry', 'mk_completion', 'fn_schema', 'payload_kwargs', 'provider_req', 'get_api_key', 'wrap_typed',
-           'unwrap_typed', 'resize_b64', 'model_prices_meta', 'infer_api_name', 'get_model_meta', 'register_model_info',
-           'get_model_info', 'effort_code', 'effort_levels', 'resolve_effort', 'get_model_pricing', 'approx_pricing',
+           'sonn', 'sonn5', 'opus46', 'opus48', 'opus', 'opus5', 'fable', 'fable5', 'fable51', 'gpt55', 'codex55',
+           'codex53spark', 'model_info_registry', 'modern_llm', 'effort_codes', 'mimo_v25_common', 'codex_pricing',
+           'sol', 'terra', 'luna', 'gpt56s', 'astra', 'Usage', 'APIRegistry', 'mk_completion', 'fn_schema',
+           'payload_kwargs', 'provider_req', 'get_api_key', 'wrap_typed', 'unwrap_typed', 'resize_b64',
+           'model_prices_meta', 'infer_api_name', 'get_model_meta', 'register_model_info', 'get_model_info',
+           'effort_code', 'effort_levels', 'resolve_effort', 'get_model_pricing', 'approx_pricing',
            'is_deepseek_peak_hour', 'price_tier', 'tier_rate']
 
 # %% ../nbs/00_types.ipynb #b4d047fd
@@ -104,7 +104,8 @@ def fn_schema(t):
 
 # %% ../nbs/00_types.ipynb #28c698fe
 def payload_kwargs(msgs, model, stream=False, system=None, max_tokens=None, temperature=None, tools=None, tool_choice=None,
-    parallel_tool_calls=None, previous_response_id=None, reasoning_effort=None, web_search_options=None, cache_idxs=None, ttl=None, stop_callables=None): pass
+    parallel_tool_calls=None, previous_response_id=None, prompt_cache_key=None, reasoning_effort=None, web_search_options=None,
+    cache_idxs=None, ttl=None, stop_callables=None): pass
 
 # %% ../nbs/00_types.ipynb #0c582256
 async def provider_req(
@@ -113,15 +114,16 @@ async def provider_req(
     body, # JSON request body
     params=None, # Query parameters
     stream=False, # Return an async generator of SSE events instead of the decoded response?
+    headers=None, # Extra headers for this request
 ):
     "POST `body` to `path` on `cli`: the decoded response, or SSE events when `stream`"
     url = cli._url(path)
     if not stream:
-        try: return dict2obj(await cli.transport.request('POST', url, params=params, json=body))
+        try: return dict2obj(await cli.transport.request('POST', url, params=params, json=body, headers=headers))
         except (httpx2.HTTPStatusError, httpx2.RequestError) as e: raise e.api_error() from e
     async def _events():
         try:
-            async for ev in cli.transport.stream('POST', url, params=params, json=body): yield dict2obj(ev)
+            async for ev in cli.transport.stream('POST', url, params=params, json=body, headers=headers): yield dict2obj(ev)
         except (httpx2.HTTPStatusError, httpx2.RequestError) as e: raise e.api_error() from e
     return _events()
 
@@ -197,11 +199,7 @@ opus48 = "claude-opus-4-8"
 opus = opus5 = "claude-opus-5"
 fable = fable5 = 'claude-fable-5'
 fable51 = 'claude-fable-5-1'
-gpt54 = "gpt-5.4"
-gpt54m = "gpt-5.4-mini"
 gpt55 = "gpt-5.5"
-codex54 = "gpt-5.4"
-codex54m = "gpt-5.4-mini"
 codex55 = "gpt-5.5"
 codex53spark = "gpt-5.3-codex-spark"
 
@@ -230,7 +228,8 @@ modern_llm = dict(supports_function_calling=True, supports_tool_choice=True, sup
 register_model_info('gpt-5.4', vendor_name='openai', base='gpt-5.4', supports_web_search=True)
 
 register_model_info('accounts/fireworks/models/deepseek-v4p1-flash', vendor_name='fireworks_ai',
-    base='accounts/fireworks/models/deepseek-v4-flash-0731', supports_vision=True, supports_image_input=True)
+    base='accounts/fireworks/models/deepseek-v4-flash-0731', supports_vision=True, supports_image_input=True,
+    input_cost_per_token=0.30e-6, cache_read_input_token_cost=0.006e-6, output_cost_per_token=1.20e-6)
 register_model_info('glm-5.3-flash', vendor_name='zai', base='glm-5.3-flash', reasoning_effort_levels=['low', 'high', 'max'])
 register_model_info('accounts/fireworks/models/glm-5p3-flash', vendor_name='fireworks_ai', base='glm-5.3-flash', base_vendor_name='zai')
 
@@ -299,7 +298,7 @@ def _flat_rates(vendor_name, model):
     info = model_info_registry[vendor_name, model]
     for k in [k for k in info if 'cost' in k and k not in codex_pricing and k != 'search_context_cost_per_query']: del info[k]
 
-for model in (codex54, codex54m, codex55):
+for model in (codex55,):
     register_model_info(model, 'codex', base=model, base_vendor_name='chatgpt', supports_web_search=True, max_input_tokens=256000, **codex_pricing)
     _flat_rates('codex', model)
 

{python_fastllm-0.0.61.dist-info → python_fastllm-0.0.63.dist-info}/METADATA

@@ -1,8 +1,8 @@
 Metadata-Version: 2.4
 Name: python-fastllm
-Version: 0.0.61
+Version: 0.0.63
 Author-email: Kerem Turgutlu <keremturgutlu@gmail.com>
-License: Apache-2.0
+License-Expression: Apache-2.0
 Project-URL: Repository, https://github.com/AnswerDotAI/fastllm
 Keywords: nbdev
 Classifier: Programming Language :: Python :: 3

python_fastllm-0.0.63.dist-info/RECORD

@@ -0,0 +1,18 @@
+fastllm/__init__.py,sha256=_yt9UDgFClBMmhhKeVz45D1Jlvpid0dbSrPSy_2bVtI,131
+fastllm/_modidx.py,sha256=iyQLdOYLOPp87AmgdFQhL7QUxXMKh-TooHd6swPFuMc,46958
+fastllm/acomplete.py,sha256=GNKQDaaDOJXzuVHozdkxnhwYFf9wZ_89UWNt7VxHLCE,11071
+fastllm/anthropic.py,sha256=N8gfoDftlH-WFVO3CED4OGTv5qDTDGBHzp4JevgOFss,19347
+fastllm/chat.py,sha256=fetuhePvXph2RnDpeQfA5SrRxgeMTmGOJA8XGXihh3s,25366
+fastllm/codex.py,sha256=r6auZ8egia83aGkWcF4132AzW4SsBvKeDwg_CxyXj9E,5808
+fastllm/gemini.py,sha256=56C3NJt8WLwx_XUWU-AElyfz_V5Zp2hIGmmkFhjFWn0,17527
+fastllm/model_prices.json,sha256=hu5XfKDhsGQylcV75uJ-i8IXjz7jDH7ifTqsS56FvAg,2367128
+fastllm/openai_chat.py,sha256=SDa8g5F5F58Li2pFHd9tSLS45f-OnI0VWvhfkP16aWM,12501
+fastllm/openai_responses.py,sha256=5WIPznQQm2DCeUo520Yqd2qjqql6EZ3UodA1gyXCiiE,15518
+fastllm/responses.py,sha256=nzB6Q90LTmJjZeSoUk36pEQQtRNcxBHkQN6NG7RlM9Y,23569
+fastllm/streaming.py,sha256=dQ7ZFc4JEtPanoKXizww91rmYXTODTisQJb1FlSfWMU,8765
+fastllm/types.py,sha256=Q1c6ZmhF_fMqE8EdZWVE50oc3SdDvD4hMuzO8MaKX_4,19368
+python_fastllm-0.0.63.dist-info/METADATA,sha256=0TjN3ymQK-UTkJ3mqlkdk-WZkOCEjwbcblurBZfWMyw,8950
+python_fastllm-0.0.63.dist-info/WHEEL,sha256=YVMoNqKzERt-wjUZwJ33xBGAwnFl-4cqbYkTtWa4itE,91
+python_fastllm-0.0.63.dist-info/entry_points.txt,sha256=WkWd04F3xVYFP5bsexaQe9Se6uelwE5ULoTtHWreO0s,133
+python_fastllm-0.0.63.dist-info/top_level.txt,sha256=F8qodL7nEGUHGmzzqfhNKCTIr1i0D6cvudOnm3z7o0Y,8
+python_fastllm-0.0.63.dist-info/RECORD,,

python_fastllm-0.0.61.dist-info/RECORD

@@ -1,18 +0,0 @@
-fastllm/__init__.py,sha256=AYKW_EYx80n3SGgwsHo7M-J6zr9komnSTd3er5PYZ8k,131
-fastllm/_modidx.py,sha256=iyQLdOYLOPp87AmgdFQhL7QUxXMKh-TooHd6swPFuMc,46958
-fastllm/acomplete.py,sha256=6qhl1RL9muhQ7ESOTq9LDS3huwUo1k2kUve_1cETruo,10881
-fastllm/anthropic.py,sha256=N8gfoDftlH-WFVO3CED4OGTv5qDTDGBHzp4JevgOFss,19347
-fastllm/chat.py,sha256=f92AOIpJipEi3GZPnTKinJvxMU2RLdu_qEKkAUjgtcU,25099
-fastllm/codex.py,sha256=r6auZ8egia83aGkWcF4132AzW4SsBvKeDwg_CxyXj9E,5808
-fastllm/gemini.py,sha256=56C3NJt8WLwx_XUWU-AElyfz_V5Zp2hIGmmkFhjFWn0,17527
-fastllm/model_prices.json,sha256=hu5XfKDhsGQylcV75uJ-i8IXjz7jDH7ifTqsS56FvAg,2367128
-fastllm/openai_chat.py,sha256=SDa8g5F5F58Li2pFHd9tSLS45f-OnI0VWvhfkP16aWM,12501
-fastllm/openai_responses.py,sha256=n0uzAULex61rXNWuzvxSBV9-qHpqSQfl-7KBNk_GO6w,15440
-fastllm/responses.py,sha256=k5D3YS8zgvUBpk_qADPtvVD9CUOWl4iej_0qvi-GuVs,23166
-fastllm/streaming.py,sha256=dQ7ZFc4JEtPanoKXizww91rmYXTODTisQJb1FlSfWMU,8765
-fastllm/types.py,sha256=RrThFMYm9mARUVbyJLiOrPkSxJtJqIe1VcvmUYzYSsY,19301
-python_fastllm-0.0.61.dist-info/METADATA,sha256=0RAXbczKkO9yehLg4q67C5ZKBxIWLKO5aSHGIvxBdJI,8939
-python_fastllm-0.0.61.dist-info/WHEEL,sha256=YVMoNqKzERt-wjUZwJ33xBGAwnFl-4cqbYkTtWa4itE,91
-python_fastllm-0.0.61.dist-info/entry_points.txt,sha256=WkWd04F3xVYFP5bsexaQe9Se6uelwE5ULoTtHWreO0s,133
-python_fastllm-0.0.61.dist-info/top_level.txt,sha256=F8qodL7nEGUHGmzzqfhNKCTIr1i0D6cvudOnm3z7o0Y,8
-python_fastllm-0.0.61.dist-info/RECORD,,