synth-ai 0.2.16__py3-none-any.whl → 0.2.17__py3-none-any.whl

examples/multi_step/configs/crafter_rl_stepwise_hosted_judge.toml

@@ -1,17 +1,10 @@
-# Crafter RL experiment – stepwise shaping with hosted judge rubrics
-#
-# This configuration extends the stepwise LoRA baseline by wiring the Synth judge
-# service so evaluation rolls combine dense step rewards with hosted rubric scoring.
-
 [algorithm]
 type = "online"
 method = "policy_gradient"
 variety = "gspo"
 
 [services]
-# Replace with the Modal URL printed by `uvx synth-ai modal-serve grpo-crafter`
 task_url = "https://YOUR-MODAL-TASK-APP.modal.run"
-# Point at the Synth backend (or compatible service) that exposes /api/judge/v1/*
 judge_url = "https://synth-backend-dev-docker.onrender.com/api"
 
 [compute]
@@ -41,7 +34,7 @@ label = "crafter-rl-stepwise-hosted-judge"
 r = 16
 alpha = 32
 dropout = 0.05
-target_modules = ["all-linear"]
+target_modules = [ "all-linear",]
 
 [rollout]
 env_name = "crafter"
@@ -50,27 +43,12 @@ episodes_per_batch = 2
 policy_name = "crafter-react"
 max_concurrent_rollouts = 8
 batches_per_step = 2
-ops = ["agent", "env"]
-
-  [rollout.env_config]
-  difficulty = "easy"
-
-    [rollout.env_config.step_rewards]
-    enabled = true
-    mode = "decision_stepwise"
-    strategy = "consistent"       # +1 for each decision that unlocks a new achievement
-    indicator_lambda = 1.0
-    step_beta = 0.0
-
-  [rollout.policy_config]
-  temperature = 0.2
-  top_p = 0.95
-  max_tokens = 512
+ops = [ "agent", "env",]
 
 [evaluation]
 instances = 16
 every_n_iters = 10
-seeds = [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15]
+seeds = [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15,]
 
 [training]
 num_epochs = 1
@@ -84,104 +62,84 @@ learning_rate = 5e-5
 log_interval = 1
 weight_sync_interval = 1
 event_rewards_kind = "unique"
-async_semaphore_max = 4  # Max concurrent rollouts in streaming pipeline
-
-# Enable dense decision rewards in the trainer to mirror env_config step rewards.
+async_semaphore_max = 4
 step_rewards_enabled = true
 step_rewards_mode = "decision_stepwise"
 step_rewards_indicator_lambda = 1.0
 step_rewards_beta = 0.0
 step_rewards_strategy = "consistent"
 
+[rubric]
+enabled = true
+
+[rollout.env_config]
+difficulty = "easy"
+
+[rollout.policy_config]
+temperature = 0.2
+top_p = 0.95
+max_tokens = 512
+
 [training.weight_sync]
 enable = true
-targets = ["policy"]
+targets = [ "policy",]
 mode = "direct"
 direct = true
 verify_every_k = 0
 
-[rubric]
-enabled = true
-model = "openai/gpt-oss-120b"
-api_base = "https://synth-backend-dev-docker.onrender.com/api/judge"
-api_key_env = "OPENAI_API_KEY"
-# Blend the hosted judge scores with environment returns inside the trainer.
 [rubric.weights]
 env = 0.2
 event = 0.4
 outcome = 0.4
 
-[rubric.event]
-# Hosted judge rubric for per-decision progress scoring.
-rubric_id = "crafter/event@v1"
-criteria = [
-  { key = "progress.unique_achievements", weight = 0.9, description = "Return 1 when this decision explicitly unlocks a brand-new Crafter achievement (inventory or status text confirms it this turn). Otherwise return 0.", aggregation = "weighted_sum" },
-  { key = "process.intent_alignment", weight = 0.1, description = "Use at most 0.3 to acknowledge tightly coupled setup that finishes the last prerequisite; keep ≤0.1 when the agent only repositions or gathers without an imminent unlock.", aggregation = "weighted_sum" },
-]
-
-[rubric.outcome]
-# Hosted judge rubric for final trajectory scoring.
-rubric_id = "crafter/outcome@v1"
-criteria = [
-  { key = "outcome.goal_completion", weight = 0.6, description = "Full credit when the agent ends with strong survival metrics and a clear crafted milestone (e.g., iron tools, furnace).", aggregation = "weighted_sum" },
-  { key = "outcome.achievement_depth", weight = 0.4, description = "Partial credit for intermediate achievements (saplings, wood/stone tools) that set up future success.", aggregation = "weighted_sum" },
-]
-
-[judge]
-type = "groq"                            # or "groq" when routing to Groq-hosted judges
+[judge.options]
+event = true
+outcome = true
+provider = "openai"
+model = "openai/gpt-oss-120b"
+rubric_id = "crafter/bundle@v1"
 timeout_s = 45
 
-  [judge.options]
-  event = true
-  outcome = true
-  provider = "openai"
-  model = "openai/gpt-oss-120b"
-  rubric_id = "crafter/bundle@v1"
-  max_concurrency = 6
-  tracks = ["process", "reasoning", "progress", "outcome"]
-
-    [judge.options.rubric_overrides]
-
-      [judge.options.rubric_overrides.event]
-      goal_text = """
-Treat each decision as a check for new Crafter achievements.
-Award the top score only when the log shows a fresh achievement unlock or an immediately verifiable deterministic completion.
-Keep otherwise useful setup actions in a narrow low band so non-achievement turns stay near zero."""
-      aggregation = "weighted_sum"
-
-        [[judge.options.rubric_overrides.event.criteria]]
-        id = "progress.unique_achievements"
-        weight = 0.9
-        scale = "binary"
-        description = "Return 1 when this decision explicitly unlocks a brand-new Crafter achievement (inventory or status text confirms it this turn). Otherwise return 0."
-
-        [[judge.options.rubric_overrides.event.criteria]]
-        id = "process.intent_alignment"
-        weight = 0.1
-        scale = "bounded"
-        description = "Use at most 0.3 to acknowledge tightly coupled setup that finishes the last prerequisite; keep ≤0.1 when the agent only repositions or gathers without an imminent unlock."
-
-      [judge.options.rubric_overrides.outcome]
-      goal_text = """
-Summarise the episode outcome in relation to Crafter’s win condition:
-survive, accumulate resources, and craft advanced tools or structures.
-Highlight notable achievements, safety failures, and preparedness for future exploration."""
-      aggregation = "weighted_sum"
-
-        [[judge.options.rubric_overrides.outcome.criteria]]
-        id = "outcome.goal_completion"
-        weight = 0.6
-        scale = "binary"
-        description = "Full credit when the agent ends with strong survival metrics and a clear crafted milestone (e.g., iron tools, furnace)."
-
-        [[judge.options.rubric_overrides.outcome.criteria]]
-        id = "outcome.achievement_depth"
-        weight = 0.4
-        scale = "bounded"
-        description = "Partial credit for intermediate achievements (saplings, wood/stone tools) that set up future success."
-
-    [judge.options.weights]
-    process = 0.05
-    reasoning = 0.15
-    progress = 0.30
-    outcome = 0.50
+[rollout.env_config.step_rewards]
+enabled = true
+mode = "decision_stepwise"
+strategy = "consistent"
+indicator_lambda = 1.0
+step_beta = 0.0
+
+[judge.options.weights]
+process = 0.05
+reasoning = 0.15
+progress = 0.3
+outcome = 0.5
+
+[judge.options.rubric_overrides.event]
+goal_text = "Treat each decision as a check for new Crafter achievements.\nAward the top score only when the log shows a fresh achievement unlock or an immediately verifiable deterministic completion.\nKeep otherwise useful setup actions in a narrow low band so non-achievement turns stay near zero."
+aggregation = "weighted_sum"
+[[judge.options.rubric_overrides.event.criteria]]
+id = "progress.unique_achievements"
+weight = 0.9
+scale = "binary"
+description = "Return 1 when this decision explicitly unlocks a brand-new Crafter achievement (inventory or status text confirms it this turn). Otherwise return 0."
+
+[[judge.options.rubric_overrides.event.criteria]]
+id = "process.intent_alignment"
+weight = 0.1
+scale = "bounded"
+description = "Use at most 0.3 to acknowledge tightly coupled setup that finishes the last prerequisite; keep ≤0.1 when the agent only repositions or gathers without an imminent unlock."
+
+[judge.options.rubric_overrides.outcome]
+goal_text = "Summarise the episode outcome in relation to Crafter’s win condition:\nsurvive, accumulate resources, and craft advanced tools or structures.\nHighlight notable achievements, safety failures, and preparedness for future exploration."
+aggregation = "weighted_sum"
+[[judge.options.rubric_overrides.outcome.criteria]]
+id = "outcome.goal_completion"
+weight = 0.6
+scale = "binary"
+description = "Full credit when the agent ends with strong survival metrics and a clear crafted milestone (e.g., iron tools, furnace)."
+
+[[judge.options.rubric_overrides.outcome.criteria]]
+id = "outcome.achievement_depth"
+weight = 0.4
+scale = "bounded"
+description = "Partial credit for intermediate achievements (saplings, wood/stone tools) that set up future success."
+

examples/multi_step/configs/crafter_rl_stepwise_shaped.toml

@@ -6,7 +6,7 @@ method = "policy_gradient"
 variety = "gspo"
 
 [services]
-# Replace with the Modal URL printed by `uvx synth-ai modal-serve grpo-crafter`
+# Replace with the Modal URL printed by `uvx synth-ai deploy --runtime modal --modal-mode serve grpo-crafter`
 task_url = "https://YOUR-MODAL-TASK-APP.modal.run"
 
 [compute]

examples/multi_step/configs/crafter_rl_stepwise_simple.toml

@@ -6,7 +6,7 @@ method = "policy_gradient"
 variety = "gspo"
 
 [services]
-# Replace with the Modal URL printed by `uvx synth-ai modal-serve grpo-crafter`
+# Replace with the Modal URL printed by `uvx synth-ai deploy --runtime modal --modal-mode serve grpo-crafter`
 task_url = "https://YOUR-MODAL-TASK-APP.modal.run"
 
 [compute]

examples/multi_step/configs/crafter_rl_stepwise_simple_NEW_FORMAT.toml

@@ -0,0 +1,105 @@
+# Crafter RL experiment – simple stepwise rewards (1 point per new achievement)
+# This config uses the NEW unified [policy] section format
+
+[algorithm]
+type = "online"
+method = "policy_gradient"
+variety = "gspo"
+
+[services]
+# Replace with the Modal URL printed by `uvx synth-ai deploy --runtime modal --modal-mode serve grpo-crafter`
+task_url = "https://YOUR-MODAL-TASK-APP.modal.run"
+
+[compute]
+gpu_type = "H200"
+gpu_count = 2
+
+  [compute.topology]  # Nested: topology is part of compute
+  type = "single_node_split"
+  gpus_for_vllm = 1
+  gpus_for_training = 1
+  gpus_for_ref = 0
+  tensor_parallel = 1
+  reference_placement = "none"  # Reference model placement
+
+[vllm]
+tensor_parallel_size = 1
+max_model_len = 8192
+
+[judge]
+enabled = false  # Set to true to enable judge/rubric scoring
+
+# Uncomment to enable judge-based reward blending:
+# enabled = true
+# timeout_s = 45
+#
+#  [judge.reward_blend]  # How to blend env/event/outcome reward sources
+#  env = 0.2
+#  event = 0.4
+#  outcome = 0.4
+#
+#  [judge.options]
+#  provider = "openai"
+#  model = "openai/gpt-oss-120b"
+#  event = true
+#  outcome = true
+#  max_concurrency = 6
+
+# NEW: Unified [policy] section - single source of truth for model and sampling
+[policy]
+model_name = "Qwen/Qwen3-4B"
+trainer_mode = "lora"
+label = "crafter-rl-stepwise-simple"
+
+# Sampling parameters for rollouts
+max_tokens = 512
+temperature = 0.6
+top_p = 0.95
+
+[rollout]
+env_name = "crafter"
+max_turns = 10
+episodes_per_batch = 4
+policy_name = "crafter-react"
+max_concurrent_rollouts = 8
+batches_per_step = 2
+ops = ["agent", "env"]
+
+[evaluation]
+instances = 10
+every_n_iters = 10
+seeds = [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]
+
+[training]
+num_epochs = 1
+iterations_per_epoch = 10
+gradient_accumulation_steps = 1
+max_accumulated_minibatch = 1
+max_turns = 10
+batch_size = 4
+group_size = 4
+learning_rate = 5e-5
+log_interval = 1
+weight_sync_interval = 1
+
+  [training.rewards]  # Nested: Reward config under training
+  step_rewards_enabled = true
+  step_rewards_mode = "decision_stepwise"
+  step_rewards_indicator_lambda = 1.0
+  step_rewards_beta = 0.0
+  step_rewards_strategy = "consistent"
+  event_rewards_kind = "unique"
+
+  [training.lora]  # Nested: LoRA config under training
+  r = 16
+  alpha = 32
+  dropout = 0.05
+  target_modules = ["all-linear"]
+
+  [training.weight_sync]
+  enable = true
+  targets = ["policy"]
+  mode = "direct"
+  direct = true
+  verify_every_k = 0
+

examples/multi_step/configs/verilog_rl_lora.toml

@@ -1,40 +1,33 @@
-# Verilog RL experiment – LoRA training on Qwen3-0.6B
-#
-# This configuration adapts the Crafter RL setup for Verilog spec-to-RTL tasks.
-# Uses the same proven pipeline but optimized for 0.6B model and Verilog domain.
-
 [algorithm]
 type = "online"
 method = "policy_gradient"
 variety = "gspo"
 
 [services]
-# Replace with the Modal URL printed by `uvx synth-ai modal-serve grpo-verilog`
 task_url = "https://synth-laboratories--grpo-verilog-task-app-fastapi-app-dev.modal.run"
-# Point at the Synth backend (or compatible service) that exposes /api/judge/v1/*
 judge_url = "https://synth-backend-dev-docker.onrender.com/api"
 
 [compute]
-gpu_type = "H200"  # ✅ 8B model needs H200 for larger context window
-gpu_count = 2      # ✅ Minimum 2x GPUs (1 for vLLM inference + 1 for training)
+gpu_type = "H200"
+gpu_count = 2
 nodes = 1
 
 [topology]
 type = "single_node_split"
-gpus_for_vllm = 1      # ✅ vLLM for inference
-gpus_for_training = 1  # ✅ Training GPU (8B LoRA fits well)
+gpus_for_vllm = 1
+gpus_for_training = 1
 gpus_for_ref = 0
 tensor_parallel = 1
 
 [vllm]
 tensor_parallel_size = 1
-max_model_len = 24576  # ✅ Increased to 24K to accommodate long Verilog prompts (16K + 8K buffer for testbenches + history)
+max_model_len = 24576
 
 [reference]
 placement = "none"
 
 [model]
-base = "Qwen/Qwen3-8B"  # ✅ 8B model for RL training with good balance of speed and capability
+base = "Qwen/Qwen3-8B"
 trainer_mode = "lora"
 label = "verilog-rl-lora-qwen8b"
 
@@ -42,38 +35,21 @@ label = "verilog-rl-lora-qwen8b"
 r = 16
 alpha = 32
 dropout = 0.05
-target_modules = ["all-linear"]
+target_modules = [ "all-linear",]
 
 [rollout]
-env_name = "verilog"  # ✅ Changed from "crafter" to "verilog"
-max_turns = 6        # ✅ More steps for compilation chains vs Crafter's 10
-episodes_per_batch = 4  # ✅ Good batch size for 8B model
+env_name = "verilog"
+max_turns = 6
+episodes_per_batch = 4
 policy_name = "verilog-designer"
 max_concurrent_rollouts = 8
 batches_per_step = 2
-ops = ["agent", "env"]
-
-  [rollout.env_config]
-  # Verilog-specific environment settings
-  difficulty = "medium"  # Can be "easy", "medium", or "hard"
-
-    [rollout.env_config.step_rewards]
-    enabled = true
-    mode = "decision_stepwise"
-    strategy = "consistent"
-    indicator_lambda = 0.5  # ✅ Reduced from Crafter (sparser rewards)
-    step_beta = 0.0
-
-  [rollout.policy_config]
-  provider = "openai"
-  model = "Qwen/Qwen3-8B"  # ✅ Use the model being trained (8B) for rollouts
-  temperature = 0.2
-  max_tokens = 4096  # ✅ Balanced for Verilog generation while leaving room for long input prompts (testbenches + history)
+ops = [ "agent", "env",]
 
 [evaluation]
 instances = 16
 every_n_iters = 10
-seeds = [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15]
+seeds = [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15,]
 
 [training]
 num_epochs = 1
@@ -81,110 +57,91 @@ iterations_per_epoch = 5
 gradient_accumulation_steps = 1
 max_accumulated_minibatch = 1
 max_turns = 15
-batch_size = 4          # ✅ Same as Crafter (works well for 8B LoRA)
+batch_size = 4
 group_size = 4
-learning_rate = 5e-5    # ✅ Same as Crafter
+learning_rate = 5e-5
 log_interval = 1
 weight_sync_interval = 1
 event_rewards_kind = "unique"
-async_semaphore_max = 20  # Max concurrent rollouts in streaming pipeline
-
-# Enable dense decision rewards in the trainer
+async_semaphore_max = 20
 step_rewards_enabled = true
 step_rewards_mode = "decision_stepwise"
-step_rewards_indicator_lambda = 0.5  # ✅ Reduced for Verilog's sparser rewards
+step_rewards_indicator_lambda = 0.5
 step_rewards_beta = 0.0
 step_rewards_strategy = "consistent"
 
+[judge]
+enabled = true
+
+[rollout.env_config]
+difficulty = "medium"
+
+[rollout.policy_config]
+provider = "openai"
+model = "Qwen/Qwen3-8B"
+temperature = 0.2
+max_tokens = 4096
+
 [training.weight_sync]
 enable = true
-targets = ["policy"]
+targets = [ "policy",]
 mode = "direct"
 direct = true
 verify_every_k = 0
 
-[rubric]
-enabled = true
+[judge.reward_blend]
+env = 0.3
+event = 0.3
+outcome = 0.4
+
+[judge.options]
+event = true
+outcome = true
+provider = "openai"
 model = "openai/gpt-oss-120b"
-api_base = "https://synth-backend-dev-docker.onrender.com/api/judge"
-api_key_env = "OPENAI_API_KEY"
+rubric_id = "verilog/bundle@v1"
+timeout_s = 45
 
-# Blend the hosted judge scores with environment returns
-[rubric.weights]
-env = 0.3     # ✅ Higher weight on env rewards for Verilog (vs Crafter's 0.2)
-event = 0.3   # ✅ Adjusted for Verilog's different reward structure
+[rollout.env_config.step_rewards]
+enabled = true
+mode = "decision_stepwise"
+strategy = "consistent"
+indicator_lambda = 0.5
+step_beta = 0.0
+
+[judge.options.weights]
+process = 0.1
+reasoning = 0.2
+progress = 0.3
 outcome = 0.4
 
-[rubric.event]
-# Verilog-specific event rubric for process efficiency
-rubric_id = "verilog/event@v1"
-criteria = [
-  { key = "process.compilation_success", weight = 0.7, description = "Return 1.0 when compilation succeeds, 0.5 for partial success, 0.0 for failure", aggregation = "weighted_sum" },
-  { key = "process.design_iterations", weight = 0.3, description = "Reward efficient design iterations without unnecessary recompilation", aggregation = "weighted_sum" },
-]
-
-[rubric.outcome]
-# Verilog-specific outcome rubric for final results
-rubric_id = "verilog/outcome@v1"
-criteria = [
-  { key = "outcome.tests_passed", weight = 0.8, description = "Full credit when all tests pass, partial for some tests", aggregation = "weighted_sum" },
-  { key = "outcome.design_quality", weight = 0.2, description = "Code quality, documentation, and design efficiency", aggregation = "weighted_sum" },
-]
-
-[judge]
-type = "groq"
-timeout_s = 45
+[judge.options.rubric_overrides.event]
+goal_text = "      Evaluate each Verilog design decision for compilation success and process efficiency.\n      High scores for successful compilation and strategic tool usage.\n      Penalize unnecessary operations and compilation failures."
+aggregation = "weighted_sum"
+[[judge.options.rubric_overrides.event.criteria]]
+id = "process.compilation_success"
+weight = 0.7
+scale = "bounded"
+description = "Return 1.0 when compilation succeeds cleanly, 0.5 for warnings, 0.0 for errors"
+
+[[judge.options.rubric_overrides.event.criteria]]
+id = "process.design_iterations"
+weight = 0.3
+scale = "bounded"
+description = "Reward efficient write→compile→simulate workflow, penalize redundant operations"
+
+[judge.options.rubric_overrides.outcome]
+goal_text = "      Evaluate the final Verilog implementation for correctness and quality.\n      High scores for working designs that pass all tests with good code quality."
+aggregation = "weighted_sum"
+[[judge.options.rubric_overrides.outcome.criteria]]
+id = "outcome.tests_passed"
+weight = 0.8
+scale = "binary"
+description = "Full credit when all tests pass, partial credit for some tests passing"
+
+[[judge.options.rubric_overrides.outcome.criteria]]
+id = "outcome.design_quality"
+weight = 0.2
+scale = "bounded"
+description = "Code clarity, proper documentation, and efficient design patterns"
 
-  [judge.options]
-  event = true
-  outcome = true
-  provider = "openai"
-  model = "openai/gpt-oss-120b"
-  rubric_id = "verilog/bundle@v1"
-  max_concurrency = 6
-  tracks = ["process", "reasoning", "progress", "outcome"]
-
-    [judge.options.rubric_overrides]
-
-      [judge.options.rubric_overrides.event]
-      goal_text = """
-      Evaluate each Verilog design decision for compilation success and process efficiency.
-      High scores for successful compilation and strategic tool usage.
-      Penalize unnecessary operations and compilation failures."""
-      aggregation = "weighted_sum"
-
-        [[judge.options.rubric_overrides.event.criteria]]
-        id = "process.compilation_success"
-        weight = 0.7
-        scale = "bounded"
-        description = "Return 1.0 when compilation succeeds cleanly, 0.5 for warnings, 0.0 for errors"
-
-        [[judge.options.rubric_overrides.event.criteria]]
-        id = "process.design_iterations"
-        weight = 0.3
-        scale = "bounded"
-        description = "Reward efficient write→compile→simulate workflow, penalize redundant operations"
-
-      [judge.options.rubric_overrides.outcome]
-      goal_text = """
-      Evaluate the final Verilog implementation for correctness and quality.
-      High scores for working designs that pass all tests with good code quality."""
-      aggregation = "weighted_sum"
-
-        [[judge.options.rubric_overrides.outcome.criteria]]
-        id = "outcome.tests_passed"
-        weight = 0.8
-        scale = "binary"
-        description = "Full credit when all tests pass, partial credit for some tests passing"
-
-        [[judge.options.rubric_overrides.outcome.criteria]]
-        id = "outcome.design_quality"
-        weight = 0.2
-        scale = "bounded"
-        description = "Code clarity, proper documentation, and efficient design patterns"
-
-    [judge.options.weights]
-    process = 0.1
-    reasoning = 0.2
-    progress = 0.3
-    outcome = 0.4

examples/qwen_coder/configs/coder_lora_30b.toml

@@ -1,11 +1,9 @@
 # Qwen3 Coder 30B LoRA SFT – all-linear adapters
 
-type = "sft"
-
 [algorithm]
 type = "offline"
 method = "sft"
-variety = "lora"
+variety = "qlora"
 
 [job]
 model = "Qwen/Qwen3-Coder-30B-A3B-Instruct"

examples/qwen_coder/configs/coder_lora_4b.toml

@@ -1,6 +1,9 @@
 # Qwen3 Coder 4B LoRA SFT – all-linear adapters
 
-type = "sft"
+[algorithm]
+type = "offline"
+method = "sft"
+variety = "qlora"
 
 [job]
 model = "Qwen/Qwen3-4B"

examples/qwen_coder/configs/coder_lora_small.toml

@@ -1,11 +1,9 @@
 # Qwen3 Coder LoRA SFT – all-linear adapters
 
-type = "sft"
-
 [algorithm]
 type = "offline"
 method = "sft"
-variety = "fft"
+variety = "qlora"
 
 [job]
 # Smallest supported Qwen3 base; replace with the smallest Coder variant when available