Spaces:
Running
Running
Upload patch_flint_free.py
Browse files- patch_flint_free.py +60 -83
patch_flint_free.py
CHANGED
|
@@ -8,12 +8,11 @@ Uses the same `openai/` path convention as Laguna S 2.1:free:
|
|
| 8 |
IMPORTANT: recursive/flint-v1.0:free has `reasoning: { mandatory: true }` on
|
| 9 |
OpenRouter. Without sending reasoning, the Recursive provider returns
|
| 10 |
400 "Inference request failed". We patch the `openai/` routing branch in
|
| 11 |
-
llm_params.py to inject `reasoning: {"effort": "low"}`
|
| 12 |
-
|
| 13 |
|
| 14 |
-
Runs AFTER patch_models.py
|
| 15 |
-
|
| 16 |
-
Idempotent.
|
| 17 |
"""
|
| 18 |
|
| 19 |
import ast
|
|
@@ -36,6 +35,7 @@ if "RECURSIVE_FLINT_V1_FREE_MODEL_ID" not in c:
|
|
| 36 |
anchors = [
|
| 37 |
'NEMOTRON_3_5_SUPER_VL_FREE_MODEL_ID = "openai/nvidia/nemotron-3.5-super-vl:free"',
|
| 38 |
'LAGUNA_S21_FREE_MODEL_ID = "openai/poolside/laguna-s-2.1:free"',
|
|
|
|
| 39 |
]
|
| 40 |
inserted = False
|
| 41 |
for anchor in anchors:
|
|
@@ -60,15 +60,14 @@ with open(AGENT_FILE) as f:
|
|
| 60 |
ac = f.read()
|
| 61 |
|
| 62 |
if "RECURSIVE_FLINT_V1_FREE_MODEL_ID" not in ac:
|
| 63 |
-
#
|
| 64 |
-
|
| 65 |
-
imp_anchor
|
|
|
|
| 66 |
if imp_anchor in ac:
|
| 67 |
ac = ac.replace(
|
| 68 |
imp_anchor,
|
| 69 |
-
imp_anchor
|
| 70 |
-
+ " NEMOTRON_3_5_SUPER_VL_FREE_MODEL_ID,\n"
|
| 71 |
-
+ " RECURSIVE_FLINT_V1_FREE_MODEL_ID,\n",
|
| 72 |
1,
|
| 73 |
)
|
| 74 |
print("OK: added RECURSIVE_FLINT_V1 import to agent.py")
|
|
@@ -77,34 +76,35 @@ if "RECURSIVE_FLINT_V1_FREE_MODEL_ID" not in ac:
|
|
| 77 |
raise SystemExit(1)
|
| 78 |
|
| 79 |
# ---------- 3. agent.py : _available_models ----------
|
| 80 |
-
# Insert
|
| 81 |
-
|
| 82 |
-
' {"id":
|
| 83 |
-
'
|
|
|
|
| 84 |
)
|
| 85 |
-
|
| 86 |
-
' {"id":
|
| 87 |
-
'"label": "recursive/flint-v1.0:free", "recommended": True},\n'
|
|
|
|
|
|
|
| 88 |
)
|
| 89 |
-
|
| 90 |
-
|
| 91 |
-
ac = ac.replace(line_anchor, new_line, 1)
|
| 92 |
print("OK: added recursive/flint-v1.0:free to _available_models()")
|
| 93 |
else:
|
| 94 |
-
|
| 95 |
alt_anchor = (
|
| 96 |
-
' {"id":
|
| 97 |
-
'"label": "inclusionai/ling-3.0-flash:free", "recommended": True},\n'
|
| 98 |
)
|
| 99 |
if alt_anchor in ac:
|
| 100 |
ac = ac.replace(
|
| 101 |
alt_anchor,
|
| 102 |
-
alt_anchor +
|
| 103 |
1,
|
| 104 |
)
|
| 105 |
print("OK: added recursive/flint-v1.0:free to _available_models() (fallback)")
|
| 106 |
else:
|
| 107 |
-
print("WARN:
|
| 108 |
|
| 109 |
with open(AGENT_FILE, "w") as f:
|
| 110 |
f.write(ac)
|
|
@@ -114,50 +114,29 @@ else:
|
|
| 114 |
print("OK: agent.py already has RECURSIVE_FLINT_V1_FREE_MODEL_ID (idempotent)")
|
| 115 |
|
| 116 |
# ---------- 4. llm_params.py : reasoning support for recursive/flint-v1.0 ----------
|
| 117 |
-
# Flint has reasoning: { mandatory: true } on OpenRouter. Without sending
|
| 118 |
-
# the Recursive provider returns 400 "Inference request failed".
|
| 119 |
# We modify the openai/ routing block to include reasoning for flint models.
|
| 120 |
-
# The reasoning parameter is passed as a top-level LiteLLM kwarg so that when
|
| 121 |
-
# using custom_llm_provider: "openrouter", LiteLLM includes it in the request
|
| 122 |
-
# body sent to https://openrouter.ai/api/v1/chat/completions.
|
| 123 |
with open(LLM_PARAMS_FILE) as f:
|
| 124 |
lp = f.read()
|
| 125 |
|
| 126 |
if "flint" not in lp.lower():
|
| 127 |
-
#
|
| 128 |
-
|
| 129 |
-
|
| 130 |
-
|
| 131 |
-
|
| 132 |
-
|
| 133 |
-
|
| 134 |
-
|
| 135 |
-
|
| 136 |
-
|
| 137 |
-
|
| 138 |
-
|
| 139 |
-
|
| 140 |
-
|
| 141 |
-
|
| 142 |
-
|
| 143 |
-
' # FIX: Strip the "openai/" prefix - OpenRouter model IDs do not include it.\n'
|
| 144 |
-
' # e.g. "openai/deepseek/deepseek-v4-flash" -> "deepseek/deepseek-v4-flash"\n'
|
| 145 |
-
' # e.g. "openai/tencent/hy3:free" -> "tencent/hy3:free"\n'
|
| 146 |
-
' if normalized_model.startswith("openai/"):\n'
|
| 147 |
-
' or_model = normalized_model[len("openai/"):]\n'
|
| 148 |
-
' params = {\n'
|
| 149 |
-
' "model": or_model,\n'
|
| 150 |
-
' "api_base": "https://openrouter.ai/api/v1",\n'
|
| 151 |
-
' "api_key": os.environ.get("OPENROUTER_API_KEY") or api_key or "",\n'
|
| 152 |
-
' }\n'
|
| 153 |
-
' # Flint requires reasoning to be explicitly enabled (mandatory on OpenRouter)\n'
|
| 154 |
-
' if "flint" in or_model.lower():\n'
|
| 155 |
-
' params["reasoning"] = {"effort": "low"}\n'
|
| 156 |
-
' return params\n'
|
| 157 |
-
)
|
| 158 |
-
|
| 159 |
-
# Version without comments (from patch_openrouter_fix.py, the final state)
|
| 160 |
-
openai_block_v2 = (
|
| 161 |
' if normalized_model.startswith("openai/"):\n'
|
| 162 |
' or_model = normalized_model[len("openai/"):]\n'
|
| 163 |
' return {\n'
|
|
@@ -167,7 +146,8 @@ if "flint" not in lp.lower():
|
|
| 167 |
' "custom_llm_provider": "openrouter",\n'
|
| 168 |
' }\n'
|
| 169 |
)
|
| 170 |
-
|
|
|
|
| 171 |
' if normalized_model.startswith("openai/"):\n'
|
| 172 |
' or_model = normalized_model[len("openai/"):]\n'
|
| 173 |
' params = {\n'
|
|
@@ -176,14 +156,14 @@ if "flint" not in lp.lower():
|
|
| 176 |
' "api_key": os.environ.get("OPENROUTER_API_KEY") or api_key or "",\n'
|
| 177 |
' "custom_llm_provider": "openrouter",\n'
|
| 178 |
' }\n'
|
| 179 |
-
' # Flint requires reasoning
|
| 180 |
' if "flint" in or_model.lower():\n'
|
| 181 |
-
' params["
|
| 182 |
' return params\n'
|
| 183 |
)
|
| 184 |
|
| 185 |
-
#
|
| 186 |
-
|
| 187 |
' if normalized_model.startswith("openai/"):\n'
|
| 188 |
' or_model = normalized_model[len("openai/"):]\n'
|
| 189 |
' return {\n'
|
|
@@ -192,7 +172,8 @@ if "flint" not in lp.lower():
|
|
| 192 |
' "api_key": os.environ.get("OPENROUTER_API_KEY") or api_key or "",\n'
|
| 193 |
' }\n'
|
| 194 |
)
|
| 195 |
-
|
|
|
|
| 196 |
' if normalized_model.startswith("openai/"):\n'
|
| 197 |
' or_model = normalized_model[len("openai/"):]\n'
|
| 198 |
' params = {\n'
|
|
@@ -200,20 +181,17 @@ if "flint" not in lp.lower():
|
|
| 200 |
' "api_base": "https://openrouter.ai/api/v1",\n'
|
| 201 |
' "api_key": os.environ.get("OPENROUTER_API_KEY") or api_key or "",\n'
|
| 202 |
' }\n'
|
| 203 |
-
' # Flint requires reasoning
|
| 204 |
' if "flint" in or_model.lower():\n'
|
| 205 |
-
' params["
|
| 206 |
' return params\n'
|
| 207 |
)
|
| 208 |
|
| 209 |
-
if
|
| 210 |
-
lp = lp.replace(
|
| 211 |
-
print("OK: patched llm_params.py openai/ block (
|
| 212 |
-
elif
|
| 213 |
-
lp = lp.replace(
|
| 214 |
-
print("OK: patched llm_params.py openai/ block (v2 custom_llm_provider) for flint reasoning")
|
| 215 |
-
elif openai_block_v1 in lp:
|
| 216 |
-
lp = lp.replace(openai_block_v1, openai_block_v1_new)
|
| 217 |
print("OK: patched llm_params.py openai/ block (v1) for flint reasoning")
|
| 218 |
else:
|
| 219 |
print("WARN: could not find openai/ routing block in llm_params.py")
|
|
@@ -222,11 +200,10 @@ if "flint" not in lp.lower():
|
|
| 222 |
if 'startswith("openai/")' in line:
|
| 223 |
print(f" Found at line {i+1}: {line.strip()}")
|
| 224 |
lines = lp.split('\n')
|
| 225 |
-
for j in range(max(0,i-2), min(len(lines), i+
|
| 226 |
print(f" {j+1}: {lines[j]}")
|
| 227 |
print("---")
|
| 228 |
print(" - Flint model will be registered but may fail with 400 'Inference request failed'")
|
| 229 |
-
print(" - The model has reasoning:{mandatory:true} on OpenRouter")
|
| 230 |
|
| 231 |
with open(LLM_PARAMS_FILE, "w") as f:
|
| 232 |
f.write(lp)
|
|
|
|
| 8 |
IMPORTANT: recursive/flint-v1.0:free has `reasoning: { mandatory: true }` on
|
| 9 |
OpenRouter. Without sending reasoning, the Recursive provider returns
|
| 10 |
400 "Inference request failed". We patch the `openai/` routing branch in
|
| 11 |
+
llm_params.py to inject `reasoning: {"effort": "low"}` via extra_body so the
|
| 12 |
+
provider accepts the request.
|
| 13 |
|
| 14 |
+
Runs AFTER patch_models.py + patch_openrouter_fix.py (which adds the
|
| 15 |
+
openai/ routing block with custom_llm_provider). Idempotent.
|
|
|
|
| 16 |
"""
|
| 17 |
|
| 18 |
import ast
|
|
|
|
| 35 |
anchors = [
|
| 36 |
'NEMOTRON_3_5_SUPER_VL_FREE_MODEL_ID = "openai/nvidia/nemotron-3.5-super-vl:free"',
|
| 37 |
'LAGUNA_S21_FREE_MODEL_ID = "openai/poolside/laguna-s-2.1:free"',
|
| 38 |
+
'LING_3_0_FLASH_FREE_MODEL_ID = "openai/inclusionai/ling-3.0-flash:free"',
|
| 39 |
]
|
| 40 |
inserted = False
|
| 41 |
for anchor in anchors:
|
|
|
|
| 60 |
ac = f.read()
|
| 61 |
|
| 62 |
if "RECURSIVE_FLINT_V1_FREE_MODEL_ID" not in ac:
|
| 63 |
+
# Find the import block and add after the last imported constant
|
| 64 |
+
imp_anchor = " SPACE_BUNNY_ALPHA_MODEL_ID,\n"
|
| 65 |
+
if imp_anchor not in ac:
|
| 66 |
+
imp_anchor = " LING_3_0_FLASH_FREE_MODEL_ID,\n"
|
| 67 |
if imp_anchor in ac:
|
| 68 |
ac = ac.replace(
|
| 69 |
imp_anchor,
|
| 70 |
+
imp_anchor + " RECURSIVE_FLINT_V1_FREE_MODEL_ID,\n",
|
|
|
|
|
|
|
| 71 |
1,
|
| 72 |
)
|
| 73 |
print("OK: added RECURSIVE_FLINT_V1 import to agent.py")
|
|
|
|
| 76 |
raise SystemExit(1)
|
| 77 |
|
| 78 |
# ---------- 3. agent.py : _available_models ----------
|
| 79 |
+
# Insert before the closing of _available_models
|
| 80 |
+
models_anchor = (
|
| 81 |
+
' {"id": "deepseek/deepseek-v4-flash-0731", "label": "DeepSeek V4 Flash 0731", "recommended": True},\n'
|
| 82 |
+
' ]\n'
|
| 83 |
+
' return models'
|
| 84 |
)
|
| 85 |
+
new_models = (
|
| 86 |
+
' {"id": "deepseek/deepseek-v4-flash-0731", "label": "DeepSeek V4 Flash 0731", "recommended": True},\n'
|
| 87 |
+
' {"id": RECURSIVE_FLINT_V1_FREE_MODEL_ID, "label": "recursive/flint-v1.0:free", "recommended": True},\n'
|
| 88 |
+
' ]\n'
|
| 89 |
+
' return models'
|
| 90 |
)
|
| 91 |
+
if models_anchor in ac:
|
| 92 |
+
ac = ac.replace(models_anchor, new_models, 1)
|
|
|
|
| 93 |
print("OK: added recursive/flint-v1.0:free to _available_models()")
|
| 94 |
else:
|
| 95 |
+
# Fallback: try the SPACE_BUNNY line as anchor
|
| 96 |
alt_anchor = (
|
| 97 |
+
' {"id": SPACE_BUNNY_ALPHA_MODEL_ID, "label": "stealth/space-bunny-alpha", "recommended": True},\n'
|
|
|
|
| 98 |
)
|
| 99 |
if alt_anchor in ac:
|
| 100 |
ac = ac.replace(
|
| 101 |
alt_anchor,
|
| 102 |
+
alt_anchor + ' {"id": RECURSIVE_FLINT_V1_FREE_MODEL_ID, "label": "recursive/flint-v1.0:free", "recommended": True},\n',
|
| 103 |
1,
|
| 104 |
)
|
| 105 |
print("OK: added recursive/flint-v1.0:free to _available_models() (fallback)")
|
| 106 |
else:
|
| 107 |
+
print("WARN: could not find _available_models anchor - model may already be present")
|
| 108 |
|
| 109 |
with open(AGENT_FILE, "w") as f:
|
| 110 |
f.write(ac)
|
|
|
|
| 114 |
print("OK: agent.py already has RECURSIVE_FLINT_V1_FREE_MODEL_ID (idempotent)")
|
| 115 |
|
| 116 |
# ---------- 4. llm_params.py : reasoning support for recursive/flint-v1.0 ----------
|
| 117 |
+
# Flint has reasoning: { mandatory: true } on OpenRouter. Without sending
|
| 118 |
+
# reasoning, the Recursive provider returns 400 "Inference request failed".
|
| 119 |
# We modify the openai/ routing block to include reasoning for flint models.
|
|
|
|
|
|
|
|
|
|
| 120 |
with open(LLM_PARAMS_FILE) as f:
|
| 121 |
lp = f.read()
|
| 122 |
|
| 123 |
if "flint" not in lp.lower():
|
| 124 |
+
# The patched llm_params.py (after patch_models.py + patch_openrouter_fix.py)
|
| 125 |
+
# has an openai/ routing block with custom_llm_provider: "openrouter".
|
| 126 |
+
# We modify it to add reasoning via extra_body for flint models.
|
| 127 |
+
|
| 128 |
+
# The block looks like:
|
| 129 |
+
# if normalized_model.startswith("openai/"):
|
| 130 |
+
# or_model = normalized_model[len("openai/"):]
|
| 131 |
+
# return {
|
| 132 |
+
# "model": or_model,
|
| 133 |
+
# "api_base": "https://openrouter.ai/api/v1",
|
| 134 |
+
# "api_key": os.environ.get("OPENROUTER_API_KEY") or api_key or "",
|
| 135 |
+
# "custom_llm_provider": "openrouter",
|
| 136 |
+
# }
|
| 137 |
+
|
| 138 |
+
# v2 (with custom_llm_provider) - the final state after all patches
|
| 139 |
+
old_block_v2 = (
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 140 |
' if normalized_model.startswith("openai/"):\n'
|
| 141 |
' or_model = normalized_model[len("openai/"):]\n'
|
| 142 |
' return {\n'
|
|
|
|
| 146 |
' "custom_llm_provider": "openrouter",\n'
|
| 147 |
' }\n'
|
| 148 |
)
|
| 149 |
+
|
| 150 |
+
new_block_v2 = (
|
| 151 |
' if normalized_model.startswith("openai/"):\n'
|
| 152 |
' or_model = normalized_model[len("openai/"):]\n'
|
| 153 |
' params = {\n'
|
|
|
|
| 156 |
' "api_key": os.environ.get("OPENROUTER_API_KEY") or api_key or "",\n'
|
| 157 |
' "custom_llm_provider": "openrouter",\n'
|
| 158 |
' }\n'
|
| 159 |
+
' # Flint requires reasoning (mandatory on OpenRouter)\n'
|
| 160 |
' if "flint" in or_model.lower():\n'
|
| 161 |
+
' params["extra_body"] = {"reasoning": {"effort": "low"}}\n'
|
| 162 |
' return params\n'
|
| 163 |
)
|
| 164 |
|
| 165 |
+
# v1 (without custom_llm_provider)
|
| 166 |
+
old_block_v1 = (
|
| 167 |
' if normalized_model.startswith("openai/"):\n'
|
| 168 |
' or_model = normalized_model[len("openai/"):]\n'
|
| 169 |
' return {\n'
|
|
|
|
| 172 |
' "api_key": os.environ.get("OPENROUTER_API_KEY") or api_key or "",\n'
|
| 173 |
' }\n'
|
| 174 |
)
|
| 175 |
+
|
| 176 |
+
new_block_v1 = (
|
| 177 |
' if normalized_model.startswith("openai/"):\n'
|
| 178 |
' or_model = normalized_model[len("openai/"):]\n'
|
| 179 |
' params = {\n'
|
|
|
|
| 181 |
' "api_base": "https://openrouter.ai/api/v1",\n'
|
| 182 |
' "api_key": os.environ.get("OPENROUTER_API_KEY") or api_key or "",\n'
|
| 183 |
' }\n'
|
| 184 |
+
' # Flint requires reasoning (mandatory on OpenRouter)\n'
|
| 185 |
' if "flint" in or_model.lower():\n'
|
| 186 |
+
' params["extra_body"] = {"reasoning": {"effort": "low"}}\n'
|
| 187 |
' return params\n'
|
| 188 |
)
|
| 189 |
|
| 190 |
+
if old_block_v2 in lp:
|
| 191 |
+
lp = lp.replace(old_block_v2, new_block_v2)
|
| 192 |
+
print("OK: patched llm_params.py openai/ block (v2 with custom_llm_provider) for flint reasoning")
|
| 193 |
+
elif old_block_v1 in lp:
|
| 194 |
+
lp = lp.replace(old_block_v1, new_block_v1)
|
|
|
|
|
|
|
|
|
|
| 195 |
print("OK: patched llm_params.py openai/ block (v1) for flint reasoning")
|
| 196 |
else:
|
| 197 |
print("WARN: could not find openai/ routing block in llm_params.py")
|
|
|
|
| 200 |
if 'startswith("openai/")' in line:
|
| 201 |
print(f" Found at line {i+1}: {line.strip()}")
|
| 202 |
lines = lp.split('\n')
|
| 203 |
+
for j in range(max(0,i-2), min(len(lines), i+10)):
|
| 204 |
print(f" {j+1}: {lines[j]}")
|
| 205 |
print("---")
|
| 206 |
print(" - Flint model will be registered but may fail with 400 'Inference request failed'")
|
|
|
|
| 207 |
|
| 208 |
with open(LLM_PARAMS_FILE, "w") as f:
|
| 209 |
f.write(lp)
|