From 9e08ab73022c9b367decb142139dee3f1e7a6e02 Mon Sep 17 00:00:00 2001 From: wizardchen Date: Wed, 13 May 2026 00:24:22 +0800 Subject: [PATCH] fix(chat): map MaxTokens to MaxCompletionTokens for GPT-5/o-series OpenAI's GPT-5 series and o-series reasoning models (o1/o3/o4-mini) no longer accept `max_tokens` and reject non-default sampling params (`temperature`, `top_p`, `frequency_penalty`, `presence_penalty`). Azure OpenAI propagates the same constraint, returning HTTP 400 with "this model is not supported MaxTokens, please use MaxCompletionTokens". Lower the compatibility shim into `RemoteAPIChat.BuildChatCompletionRequest` (the single OpenAI-protocol egress) so every internal caller keeps using `MaxTokens` uniformly: - Add `provider.IsOpenAIReasoningOrGPT5Model` with precise prefix matching (covers `gpt-5*`, `o1`/`o1-*`, `o3`/`o3-*`, `o4`/`o4-*`; rejects `olympus-1`, `openai-*`, `o3xtra`, etc.). - When the provider is `openai`/`azure_openai` and the model matches, map `MaxTokens` to `MaxCompletionTokens` (explicit `MaxCompletionTokens` wins) and skip the unsupported sampling fields; `omitempty` keeps them off the wire. - Behavior is unchanged for gpt-4o, gpt-4, and all other providers. Add unit tests for the matcher and for the request-building path (Azure gpt-5.2, OpenAI gpt-5/o1-mini/o3/o4-mini, plus negative cases). Fixes #1283 --- internal/models/chat/remote_api.go | 50 +++++++++++----- internal/models/chat/remote_api_test.go | 78 +++++++++++++++++++++++++ internal/models/provider/openai.go | 33 +++++++++++ internal/models/provider/openai_test.go | 46 +++++++++++++++ 4 files changed, 193 insertions(+), 14 deletions(-) create mode 100644 internal/models/provider/openai_test.go diff --git a/internal/models/chat/remote_api.go b/internal/models/chat/remote_api.go index d29d8c8d..b86b2774 100644 --- a/internal/models/chat/remote_api.go +++ b/internal/models/chat/remote_api.go @@ -267,21 +267,43 @@ func (c *RemoteAPIChat) BuildChatCompletionRequest(messages []Message, opts *Cha } if opts != nil { - req.Temperature = float32(opts.Temperature) - if opts.TopP > 0 { - req.TopP = float32(opts.TopP) + // OpenAI / Azure OpenAI 的推理类模型(o-series)以及 GPT-5 系列 + // 不再支持 max_tokens,必须使用 max_completion_tokens;且不支持 + // temperature / top_p / frequency_penalty / presence_penalty 的非默认值。 + // 这里做兼容处理,避免上层无差别地传 MaxTokens 导致 400 错误: + // "this model is not supported MaxTokens, please use MaxCompletionTokens" + // 参考 issue #1283。 + isOpenAIReasoning := (c.provider == provider.ProviderOpenAI || c.provider == provider.ProviderAzureOpenAI) && + provider.IsOpenAIReasoningOrGPT5Model(c.modelName) + + if !isOpenAIReasoning { + req.Temperature = float32(opts.Temperature) + if opts.TopP > 0 { + req.TopP = float32(opts.TopP) + } + if opts.FrequencyPenalty > 0 { + req.FrequencyPenalty = float32(opts.FrequencyPenalty) + } + if opts.PresencePenalty > 0 { + req.PresencePenalty = float32(opts.PresencePenalty) + } } - if opts.MaxTokens > 0 { - req.MaxTokens = opts.MaxTokens - } - if opts.MaxCompletionTokens > 0 { - req.MaxCompletionTokens = opts.MaxCompletionTokens - } - if opts.FrequencyPenalty > 0 { - req.FrequencyPenalty = float32(opts.FrequencyPenalty) - } - if opts.PresencePenalty > 0 { - req.PresencePenalty = float32(opts.PresencePenalty) + + switch { + case isOpenAIReasoning: + // 强制使用 max_completion_tokens;MaxTokens 字段保持零值,依赖 omitempty 不发送。 + if opts.MaxCompletionTokens > 0 { + req.MaxCompletionTokens = opts.MaxCompletionTokens + } else if opts.MaxTokens > 0 { + req.MaxCompletionTokens = opts.MaxTokens + } + default: + if opts.MaxTokens > 0 { + req.MaxTokens = opts.MaxTokens + } + if opts.MaxCompletionTokens > 0 { + req.MaxCompletionTokens = opts.MaxCompletionTokens + } } // 处理 Tools diff --git a/internal/models/chat/remote_api_test.go b/internal/models/chat/remote_api_test.go index ae54e9c7..b8944995 100644 --- a/internal/models/chat/remote_api_test.go +++ b/internal/models/chat/remote_api_test.go @@ -128,6 +128,84 @@ func TestBuildChatCompletionRequest_MCPToolsFormat(t *testing.T) { } } +// TestBuildChatCompletionRequest_GPT5MaxCompletionTokens 验证 GPT-5 / o-series +// 模型的 MaxTokens 自动迁移到 MaxCompletionTokens,且采样参数被剔除。 +// 见 issue #1283:Azure OpenAI 的 gpt-5 系列模型不再支持 max_tokens 字段。 +func TestBuildChatCompletionRequest_GPT5MaxCompletionTokens(t *testing.T) { + build := func(t *testing.T, providerName, modelName string) *RemoteAPIChat { + t.Helper() + c, err := NewRemoteAPIChat(&ChatConfig{ + Source: types.ModelSourceRemote, + BaseURL: "https://example.openai.azure.com", + ModelName: modelName, + APIKey: "test-key", + ModelID: modelName, + Provider: providerName, + ExtraConfig: map[string]string{ + "api_version": "2025-04-01-preview", + }, + }) + require.NoError(t, err) + return c + } + + messages := []Message{{Role: "user", Content: "test"}} + + cases := []struct { + name string + provider string + model string + shouldRewriteMaxT bool + }{ + {"AzureOpenAI gpt-5.2", "azure_openai", "gpt-5.2", true}, + {"AzureOpenAI gpt-5-mini", "azure_openai", "gpt-5-mini", true}, + {"OpenAI gpt-5", "openai", "gpt-5", true}, + {"OpenAI o1-mini", "openai", "o1-mini", true}, + {"OpenAI o3", "openai", "o3", true}, + {"OpenAI o4-mini", "openai", "o4-mini", true}, + {"OpenAI gpt-4o (unchanged)", "openai", "gpt-4o", false}, + {"AzureOpenAI gpt-4 (unchanged)", "azure_openai", "gpt-4", false}, + } + + for _, tc := range cases { + t.Run(tc.name, func(t *testing.T) { + c := build(t, tc.provider, tc.model) + opts := &ChatOptions{ + Temperature: 0.7, + TopP: 0.9, + MaxTokens: 128, + FrequencyPenalty: 0.1, + PresencePenalty: 0.2, + } + req := c.BuildChatCompletionRequest(messages, opts, false) + + if tc.shouldRewriteMaxT { + assert.Equal(t, 0, req.MaxTokens, "MaxTokens must NOT be sent for GPT-5/o-series") + assert.Equal(t, 128, req.MaxCompletionTokens, "MaxCompletionTokens should be populated from MaxTokens") + assert.EqualValues(t, 0, req.Temperature, "temperature must be omitted") + assert.EqualValues(t, 0, req.TopP, "top_p must be omitted") + assert.EqualValues(t, 0, req.FrequencyPenalty, "frequency_penalty must be omitted") + assert.EqualValues(t, 0, req.PresencePenalty, "presence_penalty must be omitted") + } else { + assert.Equal(t, 128, req.MaxTokens) + assert.Equal(t, 0, req.MaxCompletionTokens) + assert.InDelta(t, 0.7, req.Temperature, 1e-6) + } + }) + } + + t.Run("MaxCompletionTokens takes precedence over MaxTokens", func(t *testing.T) { + c := build(t, "openai", "gpt-5.2") + opts := &ChatOptions{ + MaxTokens: 128, + MaxCompletionTokens: 2048, + } + req := c.BuildChatCompletionRequest(messages, opts, false) + assert.Equal(t, 0, req.MaxTokens) + assert.Equal(t, 2048, req.MaxCompletionTokens) + }) +} + func TestBuildChatCompletionRequest_ToolChoice(t *testing.T) { chat := newTestRemoteChat(t) messages := []Message{{Role: "user", Content: "test"}} diff --git a/internal/models/provider/openai.go b/internal/models/provider/openai.go index 208e0cb9..ba0497c7 100644 --- a/internal/models/provider/openai.go +++ b/internal/models/provider/openai.go @@ -2,6 +2,7 @@ package provider import ( "fmt" + "strings" "github.com/Tencent/WeKnora/internal/types" ) @@ -51,3 +52,35 @@ func (p *OpenAIProvider) ValidateConfig(config *Config) error { } return nil } + +// IsOpenAIReasoningOrGPT5Model 判断模型是否为 OpenAI / Azure OpenAI 的 +// 推理类(o-series)或 GPT-5 系列模型。 +// +// 这些模型在 OpenAI Chat Completions API 中: +// - 不再支持 `max_tokens`,必须使用 `max_completion_tokens`; +// - 仅支持默认的 `temperature=1`、`top_p=1`,且不支持 `frequency_penalty` / +// `presence_penalty` 等采样参数(传非默认值会被拒绝)。 +// +// 参考: +// - https://platform.openai.com/docs/api-reference/chat +// - https://learn.microsoft.com/azure/ai-services/openai/how-to/reasoning +// +// 仅基于模型名做启发式匹配;对于 Azure OpenAI,因为模型名实际上是 deployment 名, +// 用户若用了自定义部署名我们无法识别,此时仍会按普通模型处理(保持原行为)。 +func IsOpenAIReasoningOrGPT5Model(modelName string) bool { + name := strings.ToLower(strings.TrimSpace(modelName)) + if name == "" { + return false + } + if strings.HasPrefix(name, "gpt-5") { + return true + } + // o1 / o1-mini / o1-preview / o3 / o3-mini / o4-mini ... + // 必须精确匹配,避免误命中 "openai-..." 之类。 + for _, prefix := range []string{"o1", "o3", "o4"} { + if name == prefix || strings.HasPrefix(name, prefix+"-") { + return true + } + } + return false +} diff --git a/internal/models/provider/openai_test.go b/internal/models/provider/openai_test.go new file mode 100644 index 00000000..3fdc4745 --- /dev/null +++ b/internal/models/provider/openai_test.go @@ -0,0 +1,46 @@ +package provider + +import "testing" + +// TestIsOpenAIReasoningOrGPT5Model 验证 GPT-5 / o-series 模型识别逻辑。 +// 见 issue #1283:这些模型必须使用 max_completion_tokens 替代 max_tokens。 +func TestIsOpenAIReasoningOrGPT5Model(t *testing.T) { + cases := []struct { + name string + input string + want bool + }{ + {"empty", "", false}, + + {"gpt-5", "gpt-5", true}, + {"gpt-5-mini", "gpt-5-mini", true}, + {"gpt-5.2", "gpt-5.2", true}, + {"gpt-5.5-pro", "gpt-5.5-pro", true}, + {"gpt-5 mixed case", "GPT-5.4-Mini", true}, + + {"o1", "o1", true}, + {"o1-mini", "o1-mini", true}, + {"o1-preview", "o1-preview", true}, + {"o3", "o3", true}, + {"o3-mini", "o3-mini", true}, + {"o4-mini", "o4-mini", true}, + + {"gpt-4", "gpt-4", false}, + {"gpt-4o", "gpt-4o", false}, + {"gpt-4o-mini", "gpt-4o-mini", false}, + {"gpt-3.5-turbo", "gpt-3.5-turbo", false}, + + {"name starting with 'o1' but not o-series", "olympus-1", false}, + {"name starting with 'openai-'", "openai-gpt-4", false}, + {"name starting with 'o3' but not o-series", "o3xtra", false}, + {"random", "qwen-max", false}, + } + for _, tc := range cases { + t.Run(tc.name, func(t *testing.T) { + got := IsOpenAIReasoningOrGPT5Model(tc.input) + if got != tc.want { + t.Errorf("IsOpenAIReasoningOrGPT5Model(%q) = %v, want %v", tc.input, got, tc.want) + } + }) + } +}