From bf2a4217271f7b679a42060f620403aa712d46b0 Mon Sep 17 00:00:00 2001 From: Devon Rifkin Date: Mon, 13 Apr 2026 14:26:15 -0700 Subject: [PATCH] gemma4: restore e2b-style nothink prompt (#15560) Gemma 4 prompts differ when thinking is disabled for different sized models: 26b/31b emit an empty thought block, while e2b/e4b do not. Before #15490, our shared Gemma 4 renderer effectively matched the e2b behavior. #15490 changed it to always emit the empty thought block, which regressed e2b/e4b nothink behavior and led to #15536 (and possibly This change restores the previous shared behavior by removing the empty trailing thought block. It also renames the checked-in upstream chat templates so the e2b and 31b fixtures are tracked separately. A follow-up will split Gemma 4 rendering by model size. Fixes: #15536 --- model/renderers/gemma4.go | 3 - model/renderers/gemma4_reference_test.go | 129 ++++--- ...jinja2 => gemma4_31b_chat_template.jinja2} | 0 .../testdata/gemma4_e2b_chat_template.jinja2 | 344 ++++++++++++++++++ 4 files changed, 429 insertions(+), 47 deletions(-) rename model/renderers/testdata/{gemma4_chat_template.jinja2 => gemma4_31b_chat_template.jinja2} (100%) create mode 100644 model/renderers/testdata/gemma4_e2b_chat_template.jinja2 diff --git a/model/renderers/gemma4.go b/model/renderers/gemma4.go index 7b186af9e..82f5fe5b1 100644 --- a/model/renderers/gemma4.go +++ b/model/renderers/gemma4.go @@ -124,9 +124,6 @@ func (r *Gemma4Renderer) Render(messages []api.Message, tools []api.Tool, thinkV // Generation prompt. if prevMessageType != "tool_response" && prevMessageType != "tool_call" { sb.WriteString("<|turn>model\n") - if !hasThink { - sb.WriteString("<|channel>thought\n") - } } return sb.String(), nil diff --git a/model/renderers/gemma4_reference_test.go b/model/renderers/gemma4_reference_test.go index 05e0517ec..5c6458ffd 100644 --- a/model/renderers/gemma4_reference_test.go +++ b/model/renderers/gemma4_reference_test.go @@ -1,14 +1,18 @@ package renderers -// TestGemma4RendererMatchesReference verifies our renderer matches the HF -// Jinja2 chat template exactly. +// TestGemma4RendererMatchesReference verifies our renderer matches the checked-in +// Gemma 4 reference template. // -// To regenerate expected values, save gemma4Jinja2Template (below) to -// gemma4_chat_template.jinja2 and run: +// Current upstream Gemma 4 chat templates differ by model size, so the checked-in +// reference intentionally uses the shared baseline without an empty generation-time +// thought channel until renderer selection is split by size. +// +// To regenerate expected values, save the E2B template to +// gemma4_e2b_chat_template.jinja2 and run: // // python3 -c " // from jinja2 import Environment; import json -// tmpl = Environment().from_string(open('gemma4_chat_template.jinja2').read()) +// tmpl = Environment().from_string(open('gemma4_e2b_chat_template.jinja2').read()) // msgs = [{'role':'user','content':'Hello'}] // print(repr(tmpl.render(messages=msgs, bos_token='', add_generation_prompt=True))) // " @@ -26,8 +30,13 @@ import ( "github.com/stretchr/testify/assert" ) -// The full Jinja2 template is committed as testdata/gemma4_chat_template.jinja2. -// Run with VERIFY_JINJA2=1 to verify expected values against the template using uv + Python. +const ( + gemma4E2BTemplate = "testdata/gemma4_e2b_chat_template.jinja2" + gemma431BTemplate = "testdata/gemma4_31b_chat_template.jinja2" +) + +// The upstream Gemma 4 chat templates are committed by size under testdata/. +// Run with VERIFY_JINJA2=1 to verify expected values against the E2B template using uv + Python. func bashRefTool() []api.Tool { return []api.Tool{{ @@ -665,7 +674,7 @@ func TestGemma4RendererMatchesReference(t *testing.T) { { name: "user_only", messages: []api.Message{{Role: "user", Content: "Hello"}}, - expected: "<|turn>user\nHello\n<|turn>model\n<|channel>thought\n", + expected: "<|turn>user\nHello\n<|turn>model\n", }, { name: "system_user", @@ -673,7 +682,7 @@ func TestGemma4RendererMatchesReference(t *testing.T) { {Role: "system", Content: "You are helpful."}, {Role: "user", Content: "Hi"}, }, - expected: "<|turn>system\nYou are helpful.\n<|turn>user\nHi\n<|turn>model\n<|channel>thought\n", + expected: "<|turn>system\nYou are helpful.\n<|turn>user\nHi\n<|turn>model\n", }, { name: "developer_user", @@ -681,13 +690,13 @@ func TestGemma4RendererMatchesReference(t *testing.T) { {Role: "developer", Content: "You are helpful."}, {Role: "user", Content: "Hi"}, }, - expected: "<|turn>system\nYou are helpful.\n<|turn>user\nHi\n<|turn>model\n<|channel>thought\n", + expected: "<|turn>system\nYou are helpful.\n<|turn>user\nHi\n<|turn>model\n", }, { name: "tools_no_system", messages: []api.Message{{Role: "user", Content: "Hi"}}, tools: bashRefTool(), - expected: "<|turn>system\n" + bashDeclRef + "\n<|turn>user\nHi\n<|turn>model\n<|channel>thought\n", + expected: "<|turn>system\n" + bashDeclRef + "\n<|turn>user\nHi\n<|turn>model\n", }, { name: "system_tools", @@ -696,7 +705,7 @@ func TestGemma4RendererMatchesReference(t *testing.T) { {Role: "user", Content: "Hi"}, }, tools: bashRefTool(), - expected: "<|turn>system\nYou are helpful." + bashDeclRef + "\n<|turn>user\nHi\n<|turn>model\n<|channel>thought\n", + expected: "<|turn>system\nYou are helpful." + bashDeclRef + "\n<|turn>user\nHi\n<|turn>model\n", }, { name: "thinking_no_system", @@ -730,6 +739,12 @@ func TestGemma4RendererMatchesReference(t *testing.T) { think: thinkTrue(), expected: "<|turn>system\n<|think|>\nYou are helpful." + bashDeclRef + "\n<|turn>user\nHi\n<|turn>model\n", }, + { + name: "thinking_explicitly_disabled", + messages: []api.Message{{Role: "user", Content: "Hi"}}, + think: thinkFalse(), + expected: "<|turn>user\nHi\n<|turn>model\n", + }, // === Message loop paths === { @@ -744,7 +759,7 @@ func TestGemma4RendererMatchesReference(t *testing.T) { "<|turn>user\nHi\n" + "<|turn>model\nHello!\n" + "<|turn>user\nMore\n" + - "<|turn>model\n<|channel>thought\n", + "<|turn>model\n", }, { // Tool call with structured args → tool response as separate <|turn>tool turn @@ -806,7 +821,7 @@ func TestGemma4RendererMatchesReference(t *testing.T) { "<|tool_response>response:bash{value:" + q + "file1.txt\nfile2.txt" + q + "}" + "Here are the files.\n" + "<|turn>user\nRead file1.txt\n" + - "<|turn>model\n<|channel>thought\n", + "<|turn>model\n", }, { // Multiple tool calls + multiple tool responses @@ -841,7 +856,7 @@ func TestGemma4RendererMatchesReference(t *testing.T) { expected: "<|turn>user\nWhat is 2+2?\n" + "<|turn>model\n4\n" + "<|turn>user\nAnd 3+3?\n" + - "<|turn>model\n<|channel>thought\n", + "<|turn>model\n", }, // === Additional edge cases ported from original tests === { @@ -899,17 +914,17 @@ func TestGemma4RendererMatchesReference(t *testing.T) { messages: []api.Message{{Role: "user", Content: "Test"}}, tools: modeTool(), expected: "<|turn>system\n" + modeDeclRef + "\n" + - "<|turn>user\nTest\n<|turn>model\n<|channel>thought\n", + "<|turn>user\nTest\n<|turn>model\n", }, { name: "unicode_content", messages: []api.Message{{Role: "user", Content: "こんにちは"}}, - expected: "<|turn>user\nこんにちは\n<|turn>model\n<|channel>thought\n", + expected: "<|turn>user\nこんにちは\n<|turn>model\n", }, { name: "newlines_in_content", messages: []api.Message{{Role: "user", Content: "Line 1\nLine 2\nLine 3"}}, - expected: "<|turn>user\nLine 1\nLine 2\nLine 3\n<|turn>model\n<|channel>thought\n", + expected: "<|turn>user\nLine 1\nLine 2\nLine 3\n<|turn>model\n", }, { // Tool response (raw JSON) followed by user message @@ -928,7 +943,7 @@ func TestGemma4RendererMatchesReference(t *testing.T) { "<|turn>model\n<|tool_call>call:get_weather{city:" + q + "Tokyo" + q + "}" + "<|tool_response>response:get_weather{value:" + q + `{"temperature": 15, "weather": "sunny"}` + q + "}" + "<|turn>user\nThanks!\n" + - "<|turn>model\n<|channel>thought\n", + "<|turn>model\n", }, // === Ordering and whitespace edge cases === { @@ -951,7 +966,7 @@ func TestGemma4RendererMatchesReference(t *testing.T) { // User content with whitespace is trimmed name: "user_content_trimmed", messages: []api.Message{{Role: "user", Content: " hello "}}, - expected: "<|turn>user\nhello\n<|turn>model\n<|channel>thought\n", + expected: "<|turn>user\nhello\n<|turn>model\n", }, { // Empty tool call arguments @@ -975,7 +990,7 @@ func TestGemma4RendererMatchesReference(t *testing.T) { messages: []api.Message{{Role: "user", Content: "Create"}}, tools: nestedTool(), expected: "<|turn>system\n" + nestedDeclRef + "\n" + - "<|turn>user\nCreate\n<|turn>model\n<|channel>thought\n", + "<|turn>user\nCreate\n<|turn>model\n", }, { // Array type in tool declaration @@ -983,7 +998,7 @@ func TestGemma4RendererMatchesReference(t *testing.T) { messages: []api.Message{{Role: "user", Content: "Batch"}}, tools: arrayTool(), expected: "<|turn>system\n" + arrayDeclRef + "\n" + - "<|turn>user\nBatch\n<|turn>model\n<|channel>thought\n", + "<|turn>user\nBatch\n<|turn>model\n", }, { // Top-level typed union follows the template's odd stringified-list form. @@ -995,8 +1010,7 @@ func TestGemma4RendererMatchesReference(t *testing.T) { <|turn>user Hi <|turn>model -<|channel>thought -`, +`, }, { // Assistant whitespace is trimmed (strip_thinking includes | trim) @@ -1009,7 +1023,7 @@ Hi expected: "<|turn>user\nHi\n" + "<|turn>model\nspaced\n" + "<|turn>user\nMore\n" + - "<|turn>model\n<|channel>thought\n", + "<|turn>model\n", }, { // Three sequential tool responses @@ -1064,7 +1078,7 @@ Hi expected: "<|turn>user\nHi\n" + "<|turn>model\nMiddleDone\n" + "<|turn>user\nMore\n" + - "<|turn>model\n<|channel>thought\n", + "<|turn>model\n", }, { // Property with no description — just type @@ -1072,7 +1086,7 @@ Hi messages: []api.Message{{Role: "user", Content: "Count"}}, tools: countTool(), expected: "<|turn>system\n" + countDeclRef + "\n" + - "<|turn>user\nCount\n<|turn>model\n<|channel>thought\n", + "<|turn>user\nCount\n<|turn>model\n", }, { // System message with leading/trailing whitespace is trimmed @@ -1082,7 +1096,7 @@ Hi {Role: "user", Content: "Hi"}, }, expected: "<|turn>system\nYou are helpful.\n" + - "<|turn>user\nHi\n<|turn>model\n<|channel>thought\n", + "<|turn>user\nHi\n<|turn>model\n", }, { // Deeply nested map in tool call arguments (3 levels) @@ -1144,7 +1158,7 @@ Hi messages: []api.Message{{Role: "user", Content: "Set"}}, tools: enumNoDescTool(), expected: "<|turn>system\n" + enumNoDescDeclRef + "\n" + - "<|turn>user\nSet\n<|turn>model\n<|channel>thought\n", + "<|turn>user\nSet\n<|turn>model\n", }, { // System message that is only whitespace (trims to empty) @@ -1154,7 +1168,7 @@ Hi {Role: "user", Content: "Hi"}, }, expected: "<|turn>system\n\n" + - "<|turn>user\nHi\n<|turn>model\n<|channel>thought\n", + "<|turn>user\nHi\n<|turn>model\n", }, { // Empty assistant content (empty string, not nil) @@ -1167,7 +1181,7 @@ Hi expected: "<|turn>user\nHi\n" + "<|turn>model\n\n" + "<|turn>user\nMore\n" + - "<|turn>model\n<|channel>thought\n", + "<|turn>model\n", }, { // Map argument with string keys (keys NOT escaped with <|"|>) @@ -1193,7 +1207,7 @@ Hi messages: []api.Message{{Role: "user", Content: "Search"}}, tools: searchTool(), expected: "<|turn>system\n" + searchDeclRef + "\n" + - "<|turn>user\nSearch\n<|turn>model\n<|channel>thought\n", + "<|turn>user\nSearch\n<|turn>model\n", }, // === Round 3 coverage gaps === @@ -1221,7 +1235,7 @@ Hi {Role: "user", Content: "Hi"}, }, expected: "<|turn>system\n\n" + - "<|turn>user\nHi\n<|turn>model\n<|channel>thought\n", + "<|turn>user\nHi\n<|turn>model\n", }, { // Nested OBJECT property with required field @@ -1229,7 +1243,7 @@ Hi messages: []api.Message{{Role: "user", Content: "Create"}}, tools: nestedRequiredTool(), expected: "<|turn>system\n" + nestedRequiredDeclRef + "\n" + - "<|turn>user\nCreate\n<|turn>model\n<|channel>thought\n", + "<|turn>user\nCreate\n<|turn>model\n", }, { // Non-integer float in tool call argument @@ -1256,7 +1270,7 @@ Hi }, expected: "<|turn>user\nHi\n" + "<|turn>model\nResult\n" + - "<|turn>model\n<|channel>thought\n", + "<|turn>model\n", }, { // Tool content with newlines and leading/trailing whitespace trimmed @@ -1280,7 +1294,7 @@ Hi messages: []api.Message{{Role: "user", Content: "Raw"}}, tools: rawTool(), expected: "<|turn>system\n" + rawDeclRef + "\n" + - "<|turn>user\nRaw\n<|turn>model\n<|channel>thought\n", + "<|turn>user\nRaw\n<|turn>model\n", }, { // Multiple required fields at top level @@ -1288,7 +1302,7 @@ Hi messages: []api.Message{{Role: "user", Content: "Move"}}, tools: moveTool(), expected: "<|turn>system\n" + moveDeclRef + "\n" + - "<|turn>user\nMove\n<|turn>model\n<|channel>thought\n", + "<|turn>user\nMove\n<|turn>model\n", }, { // Assistant content that is ONLY thinking (strips to empty) @@ -1301,7 +1315,7 @@ Hi expected: "<|turn>user\nHi\n" + "<|turn>model\n\n" + "<|turn>user\nMore\n" + - "<|turn>model\n<|channel>thought\n", + "<|turn>model\n", }, // === Round 4: final coverage gaps === @@ -1334,7 +1348,7 @@ Hi messages: []api.Message{{Role: "user", Content: "Tag"}}, tools: arrayNoItemsTool(), expected: "<|turn>system\n" + arrayNoItemsDeclRef + "\n" + - "<|turn>user\nTag\n<|turn>model\n<|channel>thought\n", + "<|turn>user\nTag\n<|turn>model\n", }, { // OBJECT property without description but with nested properties @@ -1342,7 +1356,7 @@ Hi messages: []api.Message{{Role: "user", Content: "Update"}}, tools: objectNoDescTool(), expected: "<|turn>system\n" + objectNoDescDeclRef + "\n" + - "<|turn>user\nUpdate\n<|turn>model\n<|channel>thought\n", + "<|turn>user\nUpdate\n<|turn>model\n", }, // === Round 5: coding agent patterns === @@ -1372,7 +1386,7 @@ Hi "<|tool_response>response:bash{value:" + q + q + "}" + "Done.\n" + "<|turn>user\nThanks\n" + - "<|turn>model\n<|channel>thought\n", + "<|turn>model\n", }, { // Tool call with thinking that strips to real remaining content @@ -1392,7 +1406,7 @@ Hi "<|tool_response>response:bash{value:" + q + "main.go\ngo.mod" + q + "}" + "Let me list the files.\n" + "<|turn>user\nOK\n" + - "<|turn>model\n<|channel>thought\n", + "<|turn>model\n", }, { // Argument value containing newlines (multi-line script) @@ -1720,12 +1734,35 @@ func TestGemma4RendererToolResponseWithoutNameOrIDUsesUnknown(t *testing.T) { assert.NotContains(t, got, `response:read{value:<|"|>payload<|"|>}`) } +func TestGemma4SizeTemplateFixturesDifferAtGenerationPrompt(t *testing.T) { + e2b, err := os.ReadFile(gemma4E2BTemplate) + if err != nil { + t.Fatalf("failed to read %s: %v", gemma4E2BTemplate, err) + } + + thirtyOneB, err := os.ReadFile(gemma431BTemplate) + if err != nil { + t.Fatalf("failed to read %s: %v", gemma431BTemplate, err) + } + + assert.Contains(t, string(e2b), "{{- '<|turn>model\\n' -}}") + assert.NotContains(t, string(e2b), "{{- '<|channel>thought\\n' -}}") + assert.Contains(t, string(thirtyOneB), "{{- '<|turn>model\\n' -}}") + assert.Contains(t, string(thirtyOneB), "{{- '<|channel>thought\\n' -}}") +} + // renderWithJinja2 shells out to uv + Python to render messages through the -// Jinja2 chat template. Returns the rendered string. +// E2B Jinja2 chat template. Returns the rendered string. func renderWithJinja2(t *testing.T, messages []api.Message, tools []api.Tool, think *api.ThinkValue) string { + return renderWithJinja2Template(t, gemma4E2BTemplate, messages, tools, think) +} + +// renderWithJinja2Template shells out to uv + Python to render messages through +// the named Jinja2 chat template. Returns the rendered string. +func renderWithJinja2Template(t *testing.T, templateRelPath string, messages []api.Message, tools []api.Tool, think *api.ThinkValue) string { t.Helper() - templatePath, err := filepath.Abs("testdata/gemma4_chat_template.jinja2") + templatePath, err := filepath.Abs(templateRelPath) if err != nil { t.Fatalf("failed to get template path: %v", err) } @@ -1814,3 +1851,7 @@ print(tmpl.render(**kwargs), end="") func thinkTrue() *api.ThinkValue { return &api.ThinkValue{Value: true} } + +func thinkFalse() *api.ThinkValue { + return &api.ThinkValue{Value: false} +} diff --git a/model/renderers/testdata/gemma4_chat_template.jinja2 b/model/renderers/testdata/gemma4_31b_chat_template.jinja2 similarity index 100% rename from model/renderers/testdata/gemma4_chat_template.jinja2 rename to model/renderers/testdata/gemma4_31b_chat_template.jinja2 diff --git a/model/renderers/testdata/gemma4_e2b_chat_template.jinja2 b/model/renderers/testdata/gemma4_e2b_chat_template.jinja2 new file mode 100644 index 000000000..6ec327b0b --- /dev/null +++ b/model/renderers/testdata/gemma4_e2b_chat_template.jinja2 @@ -0,0 +1,344 @@ +{%- macro format_parameters(properties, required) -%} + {%- set standard_keys = ['description', 'type', 'properties', 'required', 'nullable'] -%} + {%- set ns = namespace(found_first=false) -%} + {%- for key, value in properties | dictsort -%} + {%- set add_comma = false -%} + {%- if key not in standard_keys -%} + {%- if ns.found_first %},{% endif -%} + {%- set ns.found_first = true -%} + {{ key }}:{ + {%- if value['description'] -%} + description:<|"|>{{ value['description'] }}<|"|> + {%- set add_comma = true -%} + {%- endif -%} + {%- if value['type'] | upper == 'STRING' -%} + {%- if value['enum'] -%} + {%- if add_comma %},{%- else -%} {%- set add_comma = true -%} {% endif -%} + enum:{{ format_argument(value['enum']) }} + {%- endif -%} + {%- elif value['type'] | upper == 'ARRAY' -%} + {%- if value['items'] is mapping and value['items'] -%} + {%- if add_comma %},{%- else -%} {%- set add_comma = true -%} {% endif -%} + items:{ + {%- set ns_items = namespace(found_first=false) -%} + {%- for item_key, item_value in value['items'] | dictsort -%} + {%- if item_value is not none -%} + {%- if ns_items.found_first %},{% endif -%} + {%- set ns_items.found_first = true -%} + {%- if item_key == 'properties' -%} + properties:{ + {%- if item_value is mapping -%} + {{- format_parameters(item_value, value['items']['required'] | default([])) -}} + {%- endif -%} + } + {%- elif item_key == 'required' -%} + required:[ + {%- for req_item in item_value -%} + <|"|>{{- req_item -}}<|"|> + {%- if not loop.last %},{% endif -%} + {%- endfor -%} + ] + {%- elif item_key == 'type' -%} + {%- if item_value is string -%} + type:{{ format_argument(item_value | upper) }} + {%- else -%} + type:{{ format_argument(item_value | map('upper') | list) }} + {%- endif -%} + {%- else -%} + {{ item_key }}:{{ format_argument(item_value) }} + {%- endif -%} + {%- endif -%} + {%- endfor -%} + } + {%- endif -%} + {%- endif -%} + {%- if value['nullable'] %} + {%- if add_comma %},{%- else -%} {%- set add_comma = true -%} {% endif -%} + nullable:true + {%- endif -%} + {%- if value['type'] | upper == 'OBJECT' -%} + {%- if value['properties'] is defined and value['properties'] is mapping -%} + {%- if add_comma %},{%- else -%} {%- set add_comma = true -%} {% endif -%} + properties:{ + {{- format_parameters(value['properties'], value['required'] | default([])) -}} + } + {%- elif value is mapping -%} + {%- if add_comma %},{%- else -%} {%- set add_comma = true -%} {% endif -%} + properties:{ + {{- format_parameters(value, value['required'] | default([])) -}} + } + {%- endif -%} + {%- if value['required'] -%} + {%- if add_comma %},{%- else -%} {%- set add_comma = true -%} {% endif -%} + required:[ + {%- for item in value['required'] | default([]) -%} + <|"|>{{- item -}}<|"|> + {%- if not loop.last %},{% endif -%} + {%- endfor -%} + ] + {%- endif -%} + {%- endif -%} + {%- if add_comma %},{%- else -%} {%- set add_comma = true -%} {% endif -%} + type:<|"|>{{ value['type'] | upper }}<|"|>} + {%- endif -%} + {%- endfor -%} +{%- endmacro -%} +{%- macro format_function_declaration(tool_data) -%} + declaration:{{- tool_data['function']['name'] -}}{description:<|"|>{{- tool_data['function']['description'] -}}<|"|> + {%- set params = tool_data['function']['parameters'] -%} + {%- if params -%} + ,parameters:{ + {%- if params['properties'] -%} + properties:{ {{- format_parameters(params['properties'], params['required']) -}} }, + {%- endif -%} + {%- if params['required'] -%} + required:[ + {%- for item in params['required'] -%} + <|"|>{{- item -}}<|"|> + {{- ',' if not loop.last -}} + {%- endfor -%} + ], + {%- endif -%} + {%- if params['type'] -%} + type:<|"|>{{- params['type'] | upper -}}<|"|>} + {%- endif -%} + {%- endif -%} + {%- if 'response' in tool_data['function'] -%} + {%- set response_declaration = tool_data['function']['response'] -%} + ,response:{ + {%- if response_declaration['description'] -%} + description:<|"|>{{- response_declaration['description'] -}}<|"|>, + {%- endif -%} + {%- if response_declaration['type'] | upper == 'OBJECT' -%} + type:<|"|>{{- response_declaration['type'] | upper -}}<|"|>} + {%- endif -%} + {%- endif -%} + } +{%- endmacro -%} +{%- macro format_argument(argument, escape_keys=True) -%} + {%- if argument is string -%} + {{- '<|"|>' + argument + '<|"|>' -}} + {%- elif argument is boolean -%} + {{- 'true' if argument else 'false' -}} + {%- elif argument is mapping -%} + {{- '{' -}} + {%- set ns = namespace(found_first=false) -%} + {%- for key, value in argument | dictsort -%} + {%- if ns.found_first %},{% endif -%} + {%- set ns.found_first = true -%} + {%- if escape_keys -%} + {{- '<|"|>' + key + '<|"|>' -}} + {%- else -%} + {{- key -}} + {%- endif -%} + :{{- format_argument(value, escape_keys=escape_keys) -}} + {%- endfor -%} + {{- '}' -}} + {%- elif argument is sequence -%} + {{- '[' -}} + {%- for item in argument -%} + {{- format_argument(item, escape_keys=escape_keys) -}} + {%- if not loop.last %},{% endif -%} + {%- endfor -%} + {{- ']' -}} + {%- else -%} + {{- argument -}} + {%- endif -%} +{%- endmacro -%} +{%- macro strip_thinking(text) -%} + {%- set ns = namespace(result='') -%} + {%- for part in text.split('') -%} + {%- if '<|channel>' in part -%} + {%- set ns.result = ns.result + part.split('<|channel>')[0] -%} + {%- else -%} + {%- set ns.result = ns.result + part -%} + {%- endif -%} + {%- endfor -%} + {{- ns.result | trim -}} +{%- endmacro -%} + +{%- macro format_tool_response_block(tool_name, response) -%} + {{- '<|tool_response>' -}} + {%- if response is mapping -%} + {{- 'response:' + tool_name + '{' -}} + {%- for key, value in response | dictsort -%} + {{- key -}}:{{- format_argument(value, escape_keys=False) -}} + {%- if not loop.last %},{% endif -%} + {%- endfor -%} + {{- '}' -}} + {%- else -%} + {{- 'response:' + tool_name + '{value:' + format_argument(response, escape_keys=False) + '}' -}} + {%- endif -%} + {{- '' -}} +{%- endmacro -%} + +{%- set ns = namespace(prev_message_type=None) -%} +{%- set loop_messages = messages -%} +{{- bos_token -}} +{#- Handle System/Tool Definitions Block -#} +{%- if (enable_thinking is defined and enable_thinking) or tools or messages[0]['role'] in ['system', 'developer'] -%} + {{- '<|turn>system\n' -}} + + {#- Inject Thinking token at the very top of the FIRST system turn -#} + {%- if enable_thinking is defined and enable_thinking -%} + {{- '<|think|>\n' -}} + {%- set ns.prev_message_type = 'think' -%} + {%- endif -%} + + {%- if messages[0]['role'] in ['system', 'developer'] -%} + {{- messages[0]['content'] | trim -}} + {%- set loop_messages = messages[1:] -%} + {%- endif -%} + + {%- if tools -%} + {%- for tool in tools %} + {{- '<|tool>' -}} + {{- format_function_declaration(tool) | trim -}} + {{- '' -}} + {%- endfor %} + {%- set ns.prev_message_type = 'tool' -%} + {%- endif -%} + + {{- '\n' -}} +{%- endif %} + +{#- Pre-scan: find last user message index for reasoning guard -#} +{%- set ns_turn = namespace(last_user_idx=-1) -%} +{%- for i in range(loop_messages | length) -%} + {%- if loop_messages[i]['role'] == 'user' -%} + {%- set ns_turn.last_user_idx = i -%} + {%- endif -%} +{%- endfor -%} + +{#- Loop through messages -#} +{%- for message in loop_messages -%} + {%- if message['role'] != 'tool' -%} + {%- set ns.prev_message_type = None -%} + {%- set role = 'model' if message['role'] == 'assistant' else message['role'] -%} + {#- Detect continuation: suppress duplicate <|turn>model when previous non-tool message was also assistant -#} + {%- set prev_nt = namespace(role=None, found=false) -%} + {%- if loop.index0 > 0 -%} + {%- for j in range(loop.index0 - 1, -1, -1) -%} + {%- if not prev_nt.found -%} + {%- if loop_messages[j]['role'] != 'tool' -%} + {%- set prev_nt.role = loop_messages[j]['role'] -%} + {%- set prev_nt.found = true -%} + {%- endif -%} + {%- endif -%} + {%- endfor -%} + {%- endif -%} + {%- set continue_same_model_turn = (role == 'model' and prev_nt.role == 'assistant') -%} + {%- if not continue_same_model_turn -%} + {{- '<|turn>' + role + '\n' }} + {%- endif -%} + + {#- Render reasoning/reasoning_content as thinking channel -#} + {%- set thinking_text = message.get('reasoning') or message.get('reasoning_content') -%} + {%- if thinking_text and loop.index0 > ns_turn.last_user_idx and message.get('tool_calls') -%} + {{- '<|channel>thought\n' + thinking_text + '\n' -}} + {%- endif -%} + + {%- if message['tool_calls'] -%} + {%- for tool_call in message['tool_calls'] -%} + {%- set function = tool_call['function'] -%} + {{- '<|tool_call>call:' + function['name'] + '{' -}} + {%- if function['arguments'] is mapping -%} + {%- set ns_args = namespace(found_first=false) -%} + {%- for key, value in function['arguments'] | dictsort -%} + {%- if ns_args.found_first %},{% endif -%} + {%- set ns_args.found_first = true -%} + {{- key -}}:{{- format_argument(value, escape_keys=False) -}} + {%- endfor -%} + {%- elif function['arguments'] is string -%} + {{- function['arguments'] -}} + {%- endif -%} + {{- '}' -}} + {%- endfor -%} + {%- set ns.prev_message_type = 'tool_call' -%} + {%- endif -%} + + {%- set ns_tr_out = namespace(flag=false) -%} + {%- if message.get('tool_responses') -%} + {#- Legacy: tool_responses embedded on the assistant message (Google/Gemma native) -#} + {%- for tool_response in message['tool_responses'] -%} + {{- format_tool_response_block(tool_response['name'] | default('unknown'), tool_response['response']) -}} + {%- set ns_tr_out.flag = true -%} + {%- set ns.prev_message_type = 'tool_response' -%} + {%- endfor -%} + {%- elif message.get('tool_calls') -%} + {#- OpenAI Chat Completions: forward-scan consecutive role:tool messages -#} + {%- set ns_tool_scan = namespace(stopped=false) -%} + {%- for k in range(loop.index0 + 1, loop_messages | length) -%} + {%- if ns_tool_scan.stopped -%} + {%- elif loop_messages[k]['role'] != 'tool' -%} + {%- set ns_tool_scan.stopped = true -%} + {%- else -%} + {%- set follow = loop_messages[k] -%} + {#- Resolve tool_call_id to function name -#} + {%- set ns_tname = namespace(name=follow.get('name') | default('unknown')) -%} + {%- for tc in message['tool_calls'] -%} + {%- if tc.get('id') == follow.get('tool_call_id') -%} + {%- set ns_tname.name = tc['function']['name'] -%} + {%- endif -%} + {%- endfor -%} + {#- Handle content as string or content-parts array -#} + {%- set tool_body = follow.get('content') -%} + {%- if tool_body is string -%} + {{- format_tool_response_block(ns_tname.name, tool_body) -}} + {%- elif tool_body is sequence and tool_body is not string -%} + {%- set ns_txt = namespace(s='') -%} + {%- for part in tool_body -%} + {%- if part.get('type') == 'text' -%} + {%- set ns_txt.s = ns_txt.s + (part.get('text') | default('')) -%} + {%- endif -%} + {%- endfor -%} + {{- format_tool_response_block(ns_tname.name, ns_txt.s) -}} + {%- else -%} + {{- format_tool_response_block(ns_tname.name, tool_body) -}} + {%- endif -%} + {%- set ns_tr_out.flag = true -%} + {%- set ns.prev_message_type = 'tool_response' -%} + {%- endif -%} + {%- endfor -%} + {%- endif -%} + + {%- if message['content'] is string -%} + {%- if role == 'model' -%} + {{- strip_thinking(message['content']) -}} + {%- else -%} + {{- message['content'] | trim -}} + {%- endif -%} + {%- elif message['content'] is sequence -%} + {%- for item in message['content'] -%} + {%- if item['type'] == 'text' -%} + {%- if role == 'model' -%} + {{- strip_thinking(item['text']) -}} + {%- else -%} + {{- item['text'] | trim -}} + {%- endif -%} + {%- elif item['type'] == 'image' -%} + {{- '<|image|>' -}} + {%- set ns.prev_message_type = 'image' -%} + {%- elif item['type'] == 'audio' -%} + {{- '<|audio|>' -}} + {%- set ns.prev_message_type = 'audio' -%} + {%- elif item['type'] == 'video' -%} + {{- '<|video|>' -}} + {%- set ns.prev_message_type = 'video' -%} + {%- endif -%} + {%- endfor -%} + {%- endif -%} + + {%- if ns.prev_message_type == 'tool_call' and not ns_tr_out.flag -%} + {{- '<|tool_response>' -}} + {%- elif not (ns_tr_out.flag and not message.get('content')) -%} + {{- '\n' -}} + {%- endif -%} + {%- endif -%} +{%- endfor -%} + +{%- if add_generation_prompt -%} + {%- if ns.prev_message_type != 'tool_response' and ns.prev_message_type != 'tool_call' -%} + {{- '<|turn>model\n' -}} + {%- endif -%} +{%- endif -%}